Blog

Why AI safety breaks at the system level

Two developments in AI have started to reveal a deeper shift in how intelligent systems are built and deployed. 

One model operates behind closed doors, supporting a small group tasked with securing critical infrastructure. Another operates in the open, generating software across extended sessions with minimal supervision.

Same field. Very different philosophies.

For AI professionals, this contrast highlights a more meaningful question than model benchmarks or parameter counts: 

What kind of AI ecosystem is emerging, and how does it shape the way AI systems are designed, deployed, and trusted?

The rise of system-level risk in AI

Recent research explores how AI safety at the model level does not always translate into system-level safety in real-world deployments.

A model can demonstrate strong model alignment during evaluation, yet exhibit entirely different behaviors when embedded within LLM agents. Once connected to tools, APIs, and external environments, the model operates within a broader agentic system that introduces new dynamics.

These dynamics include:

• Multi-step reasoning across complex workflows
• Tool use and API integration within agent frameworks
• Persistent memory in AI systems across sessions
• Interaction with external and unstructured data sources

Each layer adds complexity. Each interaction expands the AI risk surface.

The result is a shift from isolated model behavior toward emergent system behavior in AI. That shift carries implications for how AI governance and safety are understood and implemented.

So why is model alignment alone not enough?

Model alignment focuses on constraining outputs within acceptable boundaries. Techniques such as reinforcement learning from human feedback (RLHF), constitutional AI, and benchmark-driven evaluation aim to shape responses toward desired behaviors.

Key factors that drive this gap include:

• Context expansion in large language models. Agents operate across extended contexts, often combining structured and unstructured data. This creates opportunities for subtle inconsistencies to influence decisions.
• Tool integration and execution risk. Access to external tools introduces operational risk. A safe response at the language level can translate into an unsafe action at the system level.
• Goal persistence in autonomous agents. AI agents maintain objectives across multiple steps. Small deviations in reasoning can compound over time, leading to outcomes that diverge from initial intent.
• Evaluation mismatch in AI systems. Many AI evaluation frameworks focus on single-turn interactions. Agent-based systems require multi-step evaluation and scenario testing to reflect real-world usage.

Together, these factors create a gap between how AI safety is measured and how AI systems behave in production.

Analytics engineering’s AI gap: Full-stack data perspective

From clean dashboards to messy intelligence systems.

AI Accelerator InstituteKshitij Aranke

The emergence of agentic complexity

Agent-based systems represent a transition from static inference toward dynamic execution. This shift introduces a new category of challenges in AI system architecture and enterprise AI deployment.

In traditional deployments, the model serves as a component within a controlled pipeline. In agentic AI systems, the model takes on a more active role, making decisions that influence future states and downstream actions.

This creates a form of operational complexity that resembles distributed systems engineering more than standalone models.

Core characteristics of agentic complexity in AI include:

• Stateful AI interactions across time
• Non-deterministic execution in LLM agents
• Feedback loops in autonomous AI systems
• Interdependencies between tools and model reasoning

These characteristics require a different approach to AI orchestration, monitoring, and control.

What this means for enterprise AI system design

As AI systems evolve, design priorities are shifting. Model performance remains important, yet AI system reliability, observability, and governance are gaining equal weight in enterprise environments.

A few principles are starting to define best practice in AI system design:

• Design for containment in AI systemsSystems benefit from clearly defined boundaries around agent capabilities. Limiting access to sensitive tools and data reduces exposure to system-level risk.
• Prioritize observability in AI workflowsDetailed logging and monitoring enable teams to understand how decisions are made across multi-step processes. This supports both debugging and AI governance frameworks.
• Structure AI workflows explicitlyBreaking tasks into defined stages improves reliability. Structured workflows guide the model through complex processes while reducing ambiguity.
• Align evaluation with real-world AI deploymentTesting frameworks need to reflect real usage conditions. Multi-step evaluation, red teaming, and adversarial testing provide more meaningful insights than static benchmarks.

These principles reflect a broader shift toward system-level thinking in AI engineering. The focus moves from optimizing individual models to managing interactions across the entire AI stack.

AI’s split future: Control vs autonomy in frontier systems

AI is splitting in two directions. One path is controlled, restricted, and security-first. The other is open, autonomous, and scaling fast. The real question isn’t which is better, it’s what this means for trust.

AI Accelerator InstituteAndrew Lovell

A new layer of responsibility in AI governance

For organizations deploying AI, this shift introduces a new layer of responsibility. AI safety can no longer be treated as a property of the model alone. It becomes a property of the entire AI system architecture.

This includes:

• How LLM agents are configured and orchestrated
• What tools and data sources AI systems can access
• How decisions are monitored, logged, and audited
• How failures in AI systems are detected and contained

This perspective aligns closely with practices in cybersecurity, risk management, and distributed systems design. It emphasizes defense in depth, continuous monitoring, and controlled deployment environments.

The path forward for agentic AI systems

The evolution of AI systems points toward a more mature phase of development. Early progress focused on expanding model capabilities and scale. The next phase focuses on integrating those capabilities into robust, production-ready AI systems.

This transition creates opportunities for teams that invest in:

• AI system architecture and orchestration
• Agent frameworks and workflow design
• AI governance and compliance

It also raises the bar for what it means to deploy enterprise AI responsibly.

Both approaches contribute to the evolving AI ecosystem.

Closing thoughts on AI system reliability

AI is entering a phase where system design defines success. Models continue to improve, yet their impact depends on how they are embedded within complex, real-world systems.

The concept of “safe models” remains important. At the same time, it represents only one layer of a broader challenge.

For AI professionals, the opportunity lies in bridging the gap between model capability and system reliability. That work defines the next frontier of AI engineering and deployment.

It also answers a question that continues to gain relevance: What makes an AI system truly safe at scale?

The difference between a developer who gets mediocre results and one who ships faster than ever comes down to one thing:

How well they have set Claude up to succeed. 

Claude Code works best as a smarter collaborator. It’s closer to onboarding a new engineer, one who needs context, structure, and clear boundaries to do their best work.

Here is how to give it exactly that…

From the onset, analytics engineering seemed like the final evolution phase of business intelligence.

It created order through modern warehouses and declarative modeling tools. Dashboards became trusted, lineage was visible, and metrics were version-controlled.

Over time, it gained traction through communities like dbt Labs, where it was established as a discipline focused on documentation, reproducibility, and testing best practices.

With the arrival of Artificial Intelligence, the focus shifted from describing the past to predicting the future.

Systems began generating new outputs and working with probabilities rather than fixed answers. Instead of deterministic SQL queries, teams began working with uncertainty. Compared to the clean, predictable nature of dashboards, AI systems feel fundamentally different.

This was the turning point.

Analytics engineering prepared me to build reliable reports. AI requires building intelligent systems. That shift demands a full-stack mindset.

Analytics engineering foundations: What we were trained to optimize

The data warehousing tradition was the foundation of analytics engineering.

We learned to prioritize clarity of structure and dimensional modelling, drawing from texts like The Data Warehouse Toolkit. The goal was consistency and trust, where the same SQL query would always produce the same result.

This determinism became the basis of stakeholder confidence. It worked because it created stability and shared understanding across teams.

However, it also introduced a set of assumptions:

• Transformations are rigid
• Answers are exact
• The world is structured

AI challenges each of these assumptions.

The mindset gap: Deterministic pipelines vs probabilistic systems

Machine learning operates on probability, while analytics engineering is built on certainty.

Research from Harvard Business Review reinforces this shift. Thomas H. Davenport and Rajeev Ronanki explain that successful AI systems deliver value within constraints, with usefulness taking priority over perfection.

This reframes what “correct” means.

Instead of asking whether something is correct, teams focus on:

• Performance improvement
• Comparison to a baseline
• Value delivered to users

As a result, fixed validation gives way to experimentation. Metrics become distributions rather than absolutes, and progress is measured iteratively. For engineers used to deterministic systems, this shift can feel unfamiliar, yet it becomes essential.

AI’s new era: Train once, infer forever in production AI

Why the future of AI systems will be driven by inference and agent workloads.

AI Accelerator InstituteAndrew Lovell

The data problem gets harder: Messy inputs, drift, and continuous quality

AI introduces a level of complexity that structured analytics rarely encounters.

Data extends beyond clean tables to include logs, images, and unstructured text, all of which require ongoing interpretation and engineering. This data also evolves over time.

In traditional analytics, issues like null values or broken schemas were visible and relatively easy to diagnose. In AI systems, challenges emerge more subtly. Models can degrade while systems appear to function as expected.

Distributions shift.

User behavior evolves.

Language changes.

Simple checks such as row counts provide limited coverage in this context.

Modern AI systems require continuous monitoring and active data management. As Bernard Marr highlights, value from AI comes from actively governed data.

Data quality becomes an ongoing responsibility.

New responsibilities: From transformation to models and MLOps

In analytics, pipelines end at insight. In AI, they extend to action.

Dashboards support human decision-making. Models automate decisions.

This shift introduces a new set of responsibilities:

• Model deployment and rollback
• Training and evaluation
• Monitoring predictions in production
• Ensuring training consistency

The lifecycle becomes continuous rather than static.

Guidance from Google formalizes this approach under MLOps, where models are treated as production systems.

Frameworks like the ML Test Score, developed by Eric Breck, provide structured ways to assess readiness and manage risk.

The risks are well documented. D. Sculley shows how quickly complexity builds in machine learning systems when pipelines are fragile or loosely defined.

Over time, shortcuts accumulate and systems become unstable.

ML systems are engineering problems, as well as data challenges.

AI’s split future: Control vs autonomy in frontier systems

AI is splitting in two directions. One path is controlled, restricted, and security-first. The other is open, autonomous, and scaling fast. The real question isn’t which is better, it’s what this means for trust.

AI Accelerator InstituteAndrew Lovell

The full-stack reality: Infrastructure, product, and human trust

As soon as models are embedded into applications, the scope expands.

Trust also becomes a defining factor.

Dashboards present verifiable numbers. Models make decisions that impact users. This introduces new expectations around transparency, bias, and accountability.

Users want explanations. Regulators expect oversight.

Trust becomes something that is designed, measured, and maintained alongside technical performance.

Conclusion

Analytics engineering provided strong foundations in lineage, reproducibility, testing, and discipline.

AI builds on these foundations while introducing uncertainty, continuous change, and new system-level challenges.

The boundaries between engineering, analytics, and product continue to converge. Data professionals increasingly think across the full stack, from data models to real-world impact.

The goal is to extend analytics engineering.

From clean dashboards to intelligent systems. From static pipelines to adaptive ones.

This is the shift AI demands, and it highlights the gap that analytics engineering alone did not fully address.

References

• Eric Breck, Polyzotis, N., Roy, S., Whang, S., & Zinkevich, M. (2017). The ML Test Score: A Rubric for ML Production Readiness.
• Thomas H. Davenport, & Rajeev Ronanki (2018). Artificial Intelligence for the Real World. Harvard Business Review.
• Google (2020). MLOps: Continuous Delivery and Automation Pipelines in Machine Learning.
• Kimball, R., & Ross, M. (2013). The Data Warehouse Toolkit. Wiley.
• Martin Kleppmann (2017). Designing Data-Intensive Applications. O’Reilly Media.
• Bernard Marr (2021). Data Strategy: How to Profit from a World of Big Data, Analytics and AI. Kogan Page.
• D. Sculley et al. (2015). Hidden Technical Debt in Machine Learning Systems. In Neural Information Processing Systems Proceedings.

Over the past several years, the AI industry has focused heavily on training increasingly large models.

Discussions about artificial intelligence often center around massive GPU clusters, trillion-parameter architectures, and the enormous computational resources required to train modern systems. 

Training has become the most visible symbol of progress in machine learning, and it often dominates headlines across the technology industry.

However, once a model is trained, the real operational challenge begins. Every A-powered product relies on inference and the process of running a trained model to generate predictions or responses. 

As organizations move from experimentation to production environments, it becomes clearer that the long-term engineering challenge of AI is not only training models, but running them efficiently at scale.

The shift from training to operational AI

Training a modern model requires substantial computing resources, but for most organizations, it is not a daily activity. A model may be trained or fine-tuned periodically and then deployed to support applications that operate continuously.

Once deployed, the same model may serve thousands or millions of requests every day across multiple systems.

This changes how companies must think about AI infrastructure. Training represents a large but relatively short-lived workload, while inference becomes an ongoing operational workload that grows with usage.

As AI capabilities become embedded in enterprise applications, the number of inference calls increases rapidly. Over time, the cost and engineering complexity of running models in production can exceed the original cost of training them.

AI’s split future: Control vs autonomy in frontier systems

AI is splitting in two directions. One path is controlled, restricted, and security-first. The other is open, autonomous, and scaling fast. The real question isn’t which is better, it’s what this means for trust.

AI Accelerator InstituteAndrew Lovell

Inference as the operational core of AI systems

Every inference request consumes compute resources.

When a user sends a prompt to a language model, the system processes the input tokens and generates output tokens step by step.

Large language models generate responses sequentially, which means the model remains active throughout the entire generation process, continuing to use GPU memory and compute resources.

At scale, these operations become significant. Enterprise copilots, automated support systems, and AI-powered search tools may process millions of prompts each day.

The infrastructure supporting these systems must manage latency, GPU utilization, and memory constraints while maintaining predictable performance.

As organizations expand their AI deployments, the focus naturally shifts toward improving inference efficiency.

The architecture of enterprise AI platforms

The second layer prepares models for production through optimization techniques such as quantization, distillation, and parameter-efficient fine-tuning.

The final layers focus on inference infrastructure and applications where models are served through scalable APIs and integrated into products. 

In many production environments, the most complex engineering challenges occur in these later stages, where models must operate reliably under real workloads.

The machine learning paradox: Can AI actually learn?

We call it machine learning. But do machines actually learn? Today’s AI systems train, optimize, and scale, but real learning is something else entirely. The distinction matters more than the industry wants to admit.

AI Accelerator InstituteAndrew Lovell

Scaling inference for large language models

Running large language models efficiently requires several optimization techniques.

Quantization reduces the numerical precision of model weights, which allows models to run faster and consume less memory. Distillation allows smaller models to replicate the behavior of larger models for specific tasks, which can significantly reduce compute requirements.

Infrastructure-level improvements are also important. Continuous batching allows multiple requests to be processed together, which increases hardware utilization. 

Techniques such as KV cache reuse and speculative decoding improve token generation throughput and reduce latency.

These optimizations make it possible to run large models in production systems where both cost and performance matter.

Modern infrastructure for large-scale inference

As AI adoption grows, new infrastructure patterns are emerging to support inference workloads. One approach is server-less inference, where compute resources automatically scale based on demand.

Instead of maintaining GPU clusters that run continuously, the system can allocate resources dynamically as requests arrive, improving overall utilization.

Another important development is GPU sharing and multi-model serving. Instead of dedicating a GPU to a single model, modern inference platforms allow multiple models to run on the same hardware and schedule requests dynamically.

Techniques such as request batching and model multiplexing further improve efficiency by enabling the system to support many workloads without continuously expanding infrastructure.

Agents and the amplification of inference workloads

A major change in AI applications is the rise of agent-based systems. Traditional AI applications typically generate a single response to a user request. Agent systems behave differently because they perform multi-step reasoning before producing a final result.

An agent may break down a task into smaller steps, retrieve information from external systems, and generate several intermediate prompts during the process. Each step usually requires another model inference. 

As a result, a single user request may trigger many model executions instead of just one.

Agent-driven workflows, therefore, amplify the amount of inference performed by the system and increase the demand on the underlying infrastructure.

Infrastructure implications of agent workloads

To support these workloads efficiently, infrastructure must manage high volumes of model calls while maintaining predictable performance.

Techniques such as model routing, efficient batching, GPU sharing, and dynamic scaling become even more important when agent workflows operate at scale. 

As organizations adopt agent-driven automation, the importance of efficient inference infrastructure continues to grow.

Designing systems with inference efficiency in mind

As organizations gain experience with production AI deployments, many teams are beginning to design architectures that prioritize inference efficiency from the outset.

Instead of relying on a single large model, systems may route simple tasks to smaller models and reserve larger models for more complex reasoning tasks.

Other design strategies include streaming responses so users can see results as they are generated, and dynamically scaling infrastructure based on real-time demand. Efficient scheduling and GPU sharing can further improve hardware utilization and reduce operational costs.

These approaches help ensure that both language model applications and agent-driven workflows can operate reliably at scale.

The story of Sora: What it reveals about building real-world AI

After ChatGPT’s breakthrough, the race to define the next frontier of generative AI accelerated. One of the most talked-about innovations was OpenAI’s Sora, a text-to-video AI model that promised to transform digital content creation.

AI Accelerator InstituteAndrew Lovell

The future of AI infrastructure

The broader technology ecosystem is beginning to adapt to the growing importance of inference workloads.

Hardware vendors are developing accelerators optimized specifically for inference performance, while cloud platforms are introducing systems designed for large-scale model serving.

As agent-based applications become more common, the number of inference requests will continue to increase. 

Conclusion

Artificial intelligence is entering a new stage of maturity. Early progress focused on training large models and demonstrating the capabilities of modern machine learning systems. These breakthroughs established the foundation for the rapid expansion of AI across industries.

As AI becomes embedded in real applications, the focus is shifting toward how these systems operate in production environments. Inference now represents the core workload that powers both language models and agent-driven systems. 

Organizations that design infrastructure optimized for efficient inference will be best positioned to support the next generation of intelligent applications. In the long run, training happens occasionally, but inference and agent execution happen continuously.

Have you ever wondered just how long it takes to bring a new medicine to market? For those of us working in the pharmaceutical industry, the answer is clear: it can take decades and cost billions. As a computer scientist leading the AI team at AstraZeneca, I’ve seen firsthand just how complicated the drug discovery process is.

Trust me, there’s a lot of work that goes into developing a drug. But the real challenge lies in making the process more efficient and accessible, which is where generative AI comes in.

In this article, I’ll walk you through the critical role that AI plays in accelerating drug discovery, particularly in areas like medical imaging, predicting patient outcomes, and creating synthetic data to address data scarcity. While AI has incredible potential, it’s not without its challenges. From building trust with pathologists to navigating regulatory requirements, there’s a lot to consider. 

So, let’s dive into how AI is reshaping the way we approach drug discovery and development – and why it matters to the future of healthcare.

Generative AI’s role in enhancing medical imaging

One of the most powerful applications of generative AI in drug development is its ability to analyze medical images – a process that’s essential for diagnosing diseases like cancer, which can be difficult to detect early on.

In the world of pathology, we’re increasingly moving away from using traditional microscopes, and using digital images instead. With digitized tissue biopsies, we now have access to incredibly detailed scans that show every single cell. The sheer volume of this data – sometimes containing over 100,000 x 100,000 pixels – makes it almost impossible for human pathologists to analyze every single detail, but AI can handle this level of complexity.

At AstraZeneca, we’ve been using generative AI to help analyze these vast datasets. One of the most exciting developments is in medical imaging, where AI models can quickly identify cancerous cells, segment different areas of tissue, and even predict patient outcomes. 

For example, AI can help predict whether a patient will respond to a specific drug – information that’s invaluable for companies like ours, as we work to develop treatments that will provide real, tangible benefits for patients.

In my work, we leverage powerful AI techniques such as variational autoencoders and generative adversarial networks (GANs) to build these models. These AI techniques can help us learn from medical images and generate synthetic data that can be used to train AI models more effectively.

Computer Vision in Healthcare: Download the eBook today

Unlock the mystery of the innovative intersection of technology and medicine with our latest eBook, Computer Vision in Healthcare.

AI Accelerator InstituteMarisa Garanhel

Overcoming data scarcity with synthetic data

In healthcare, one of the biggest challenges we face is data scarcity. High-quality, labeled medical data is incredibly difficult to obtain, and privacy concerns mean that it’s often not feasible to collect large amounts of patient data. This scarcity of data can be a significant barrier to building reliable AI models.

One of the ways we’re addressing this challenge is by using synthetic data. With generative AI, we can create realistic medical data that mirrors the patterns seen in real-world datasets. This synthetic data can be used to train AI models, allowing us to overcome some of the limitations posed by data scarcity.

In addition to providing data for training, synthetic data can help address another key issue in medical imaging: heterogeneity. Data collected from different hospitals or medical centers may differ in terms of image quality, scanning techniques, or even how the tissue samples are prepared. These differences can make it difficult for AI models to generalize and perform well across datasets from different sources.

To solve this, we’ve explored using generative AI to standardize the data. For example, if we train a model using data from one hospital, it may not work as well when applied to data from a different hospital. 

By generating synthetic data that mimics the characteristics of different hospitals, we can ensure that our models work reliably across a wide range of real-world scenarios.

The impact and challenges of generative AI in healthcare

Examining generative AI in healthcare: innovations, consumer views, and pressing ethical considerations.

AI Accelerator InstituteMarisa Garanhel

The role visual language models play in drug development

A particularly exciting area of research for us has been the use of visual language models. These models combine both visual data (such as medical images) and textual data (such as pathology reports) to create more comprehensive insights.

Imagine this: a pathologist receives a biopsy scan and needs to write a report that summarizes the findings. Instead of manually writing out every detail, a visual language model could analyze the image and generate an automatic report, outlining exactly what’s happening in the scan – what areas are abnormal, where cancer is present, and whether any other pathologies are visible.

In a project we worked on, we analyzed pathology reports from 14 hospitals across the UK, all related to chronic kidney disease. These reports were highly heterogeneous, with different hospitals using different formats, and the text was written in varying styles. 

Using AI, we harmonized these reports into a standardized format, extracting only the most relevant information. This multi-modal approach – combining text and images – showed just how powerful generative AI can be in streamlining the process of medical data analysis.

AI for health & networking: Christie Mealo’s tech impact

Christie Mealo discusses AI in health, her AI networking app, and her role in Philly’s tech scene, highlighting generative AI’s impact.

AI Accelerator InstituteChristie Mealo

Building trust with pathologists and medical experts

One of the biggest challenges we face in applying AI to healthcare is building trust with the medical experts who use our models. 

Pathologists, after all, have spent years developing their expertise and have a deep understanding of the nuances of disease detection. Convincing them that AI can be a useful tool – and not a replacement for their expertise – requires careful consideration and collaboration.

From the beginning, we’ve worked closely with pathologists to ensure that the AI models we develop address their real-world needs. We don’t simply create a model and then hand it off to the pathologist; instead, we sit down with them, understand their challenges, and develop the model together. This collaborative approach ensures that the AI we build is both useful and trustworthy.

One example of this collaboration is our work on prostate cancer detection. The challenge with prostate cancer is that some cancerous glands are so small that even experienced pathologists can miss them. But AI can spot these tiny glands with impressive accuracy.

When we showed our model’s results to pathologists, they were impressed by how effectively the AI was able to detect these hard-to-find glands. This success then helped build their trust in the technology.

Regulating artificial intelligence: The bigger picture

The article discusses the challenges of AI regulation, economic impact, and governance, with a focus on the UK’s evolving legal approach.

AI Accelerator InstituteAna Simion

Ensuring accountability and regulatory compliance

While AI can significantly enhance the drug discovery process, it’s important to remember that pathologists remain accountable for the final medical decisions. AI is a tool that supports their work, but it’s not a replacement for their judgment.

At AstraZeneca, we are deeply committed to ensuring that all of our AI applications comply with regulatory standards. The healthcare industry is heavily regulated, and we take compliance seriously. Every model we develop must meet rigorous standards before it can be deployed in clinical settings.

We also recognize that developing AI models for healthcare involves an ethical responsibility. These models must be transparent, explainable, and free from bias. It’s not enough for the model to work; it must also work fairly and reliably across all patient populations.

Conclusion

So, while AI is an incredibly powerful tool, it’s essential to remember that pathologists, clinicians, and researchers will always remain at the heart of drug discovery. AI is here to support their work, not replace it. 

As we continue to advance in our use of AI, it’s exciting to think about the possibilities for improving patient outcomes and accelerating the development of new medicines. But, at the end of the day, the human touch – expert judgment, empathy, and care – will always be paramount in healthcare.

At AstraZeneca, we’re excited to be at the forefront of these changes, and we look forward to continuing our work with AI to make a meaningful difference in the lives of patients around the world.

This article comes from Dr Nikolay Burlutskiy’s talk at our London 2024 Generative AI Summit. Check out his full presentation and the wealth of OnDemand resources waiting for you.

Nikolay was the former Director of ML and AI at AstraZeneca at the time of this talk. He’s currently the Senior Manager of GenAI Platforms at Mars.

The Complier is the hardest boss to reach in the extraction shooter Marathon. To even have the chance to fight it, you need to have cleared six vaults—increasingly elaborate puzzle rooms—in the Cryo Archive, Marathon’s end game map. To even get the chance to enter each of those vaults, you need to obtain a key for each. To even get a chance to get one of those keys, you need to kill another set of bosses or find them in dangerous runs of another map. And if you do find a key, or you bring one into Cryo Archive to use, another team of players may simply kill you and take it from you. 

Or, you could pay a random guy on eBay to kill the Compiler for you. 

“Too busy with life? Want to hop on after a long day with a vault full of loot? Look no further!,” the description for a listing on eBay says. The listing itself is advertising a “Cryo Archive Compiler Kill.”

“Since the old Destiny loot cave days, I have loved helping players get the most out of their enjoyment with the game. Whether you want lots of loot, a higher rank, or a fun group of people to play with, my goal is simple: help you get results without wasting time,” it adds.

Paid boosting in video games is, obviously, not new. For years players without enough time to do it themselves have paid other people real money to grind Call of Duty experience for them, get to a certain rank in World of Warcraft, or obtain specific loot in Arc Raiders. But I found the Compiler kill offer especially jarring because it is something that requires so much time and skill from the person offering the boost. Killing, even getting to, the Compiler is not a mindless grind. You have to play a lot of Marathon to get there, and be good at the game. That, and personally it is a goal Emanuel, Matthew, and myself are slowly working towards, because that slow, painful progress is so satisfying to do yourself. 

One attraction of killing the Compiler is that you get a unique character skin after doing so, something that in the know players will definitely notice you flaunting. There is also a chance to get the Biotoxic Disinjector weapon as a reward. This is a ludicrous gun that shoots both slime and grenades, and Bungie already had to lower its power once. If you want one Biotoxic Disinjector, the booster is charging $200. If you want three, you need to cough up $400. If you’re happy with just the kill itself, it costs $125. According to the listing, 15 people have paid for this particular service.

The eBay listing says buyers can have the booster play on the customer’s account, or “You play with us (Me and one more good player) *More expensive.” They also let you pay and play with another person of your choosing, but keep it hidden from them you’re paying for a boost, if you want to add some friendship deception in there too. 

I noticed at least one listing advertising a similar Compiler kill service has been removed from eBay. Bungie, Marathon’s developer, did not respond to a request for comment, and I specifically asked Bungie if these boost services violate its rules.

Residents of an Atlanta suburb have been rocked by the revelation that sales employees at Flock have been accessing sensitive cameras in the town to demonstrate the company’s surveillance technology to police departments around the country. The cameras accessed have included surveillance tech in a children’s gymnastics room, a playground, a school, a Jewish community center, and a pool.

Flock has taken issue with the way that residents and activists have characterized the access but confirmed that the camera access did happen as part of its sales demonstrations. A blog post by Jason Hunyar, a Dunwoody, Georgia resident who learned about Flock accessing the city’s cameras by obtaining Flock access logs via a public records request is called “Why Are Flock Employees Watching Our Children?” 

Flock has pushed back against this characterization on social media, in a blog post, at city council meetings, and in a statement to 404 Media: “The city of Dunwoody is one city in our demo partner program,” a Flock spokesperson told 404 Media. “The cities involved in this program have authorized select Flock employees to demonstrate new products and features as we develop them in partnership with the city. Moreover, select engineers can access accounts with customer permission to debug or fix any issues that may arise. No one is spying on children in parks, as the substack incorrectly asserts.” 

Flock also argued that it is more transparent than any other surveillance company because it creates these access logs at all, and they can be obtained using public records requests. “Also, I must state the irony of the situation. We’re one of the few technology companies in this space dedicated to radical transparency […] I understand the concern from the resident, but it is unequivocally false to assert that Flock, or the police, or city officials are doing anything other than using technology to stop major crimes in the city.”

The records Hunyar obtained, however, show that some of the cameras that were accessed were in sensitive locations, including the pool at the Marcus Jewish Community Center of Atlanta (in Dunwoody), the children’s gymnastics room at MJCCA, and several fitness centers and studios. The access logs obtained by Hunyar show at the very least how expansive Flock’s surveillance systems can be in a single city, encompassing not just cameras purchased by the city but also cameras purchased by private businesses. 

After Hunyar wrote about what he found, Flock has agreed to stop using Dunwoody’s cameras to demonstrate its product. Flock’s FAQ page states that “Flock customers own their data” and “Flock will not share, sell, or access your data.” It also states “nobody from Flock Safety is accessing or monitoring your footage.” Flock also published a blog post that notes “one of the benefits communities value most about Flock technology is the ability for law enforcement to directly access privately owned cameras, if and only if the organization allows them to, for crime-solving and security purposes.” 

“Fair questions have been asked about conducting demos on cameras in sensitive locations when doing this very critical testing in the real-world. Last week, in the City of Dunwoody, questions were raised about a demo conducted as part of authorized activity approved under the city’s demo partner agreement, on cameras at a local Jewish Community Center. Although the camera was only viewed during a routine demo, we understand that this is a sensitive location for many. We have therefore determined that employees will be trained to only conduct demos in more public locations, like retail parking lots,” Flock wrote in the blog. “Accusing someone of spying on children is not a policy disagreement; it is a life-altering allegation. Claims of inappropriate conduct by our employees are false. The employees being named online are well-intentioned employees who accessed a camera network with the city’s explicit permission, as part of their job. They are now being called predators for it.”

Japan’s Minister of Defense Shinjirō Koizumi posed with a cardboard drone on Monday during a meeting with drone manufacturer AirKamuy. The AirKamuy 150 is a cheap pre-fab cardboard drone meant to die on the battlefield and it comes shipped in a flatpack like an IKEA shelf.

According to Koizumi, Japan’s military has already begun to use the cardboard drone. “The Japan Maritime Self-Defense Force is already utilizing them as targets,” he said in a post on X. “In aiming to become the Self-Defense Forces that makes the most extensive use of unmanned assets, including drones, in the world, strengthening collaboration with startups enthusiastic about the defense sector is indispensable.”

Update 4:30 PM EDT: On a popular listserv for academics, many of whom are attending RightsCon, a board member of Access Now wrote “I am told I can leak that RightsCon has been canceled. Message from [Access Now] following shortly” in a thread about what attendees were planning on doing. And in an email, AccessNow wrote: “It is with heavy hearts that we share: RightsCon will not proceed in Zambia or online. We understand this news is deeply upsetting for our community and while we know everyone has questions, our goal right now is to notify you of the event’s status because many of you have imminent travel plans. We do not recommend registered participants travel to Lusaka for RightsCon.

Over the last 48 hours we have experienced an overwhelming surge of support from civil society, government representatives, sponsors, and our community as a whole. For this, we wholeheartedly thank you. We’ll communicate more information soon.” The original article continues below.

Days before thousands of researchers, academics, and human rights experts were set to convene in Lusaka, Zambia, the government of Zambia announced it was postponing RightsCon, one the largest and most important digital human rights conferences in the world. The announcement, which came as some participants and speakers were already en route to the conference, has sown confusion and chaos in the academic community. 

Minister of Technology and Science Felix Mutati first announced the postponement on April 28, saying that Zambia needed more time to ensure the conference “fully [aligns] with national procedures, diplomatic protocols, and the broader objective of fostering a balanced and consensus-driven platform for dialogue.” 

“In particular, certain invited speakers and participants remain subject to pending administrative and security clearances, which have not yet been concluded,” he added, according to the Lusaka Times.

Customs and Border Protection (CBP) plans to spend hundreds of millions of dollars to expand its fleet of high-powered surveillance drones, and other parts of the Department of the Homeland Security (DHS) may buy their own Predator-style drones, according to recently published procurement records.

The news shows DHS’s continued investment in drone surveillance technology, and how use of large scale drones could expand to other parts of the umbrella agency.