SAFETY

→ View original post on X — @jiqizhixin

AI Agents Collaborating to Commit Social Media Fraud

By

@jiqizhixin

–

19 February 2026 18h51

What happens when AI agents start working together to scam people on social media? Researchers from Shanghai Jiao Tong University, Shanghai AI Lab, and Beihang University have introduced MultiAgentFraudBench to investigate this emerging threat. They simulated how groups of LLM

→ View original post on X — @fogoros

Industrial Robot Security: Access Control and Worker Safety

By

@fogoros

–

19 February 2026 17h15

In an industrial setting, a security breach means more than just leaked data. If a robot's controls are compromised, it can be programmed to move unpredictably, potentially injuring workers or destroying assets. Proper access control and governance are fundamental to worker… pic.twitter.com/pR9NGQuGRr
— Lucian Fogoros (@fogoros) 19 février 2026

In an industrial setting, a security breach means more than just leaked data. If a robot's controls are compromised, it can be programmed to move unpredictably, potentially injuring workers or destroying assets. Proper access control and governance are fundamental to worker

→ View original post on X — @dair_ai

Agent Memory Benchmarks Don’t Predict Real-World Performance

By

@dair_ai

–

19 February 2026 15h28

Agent memory benchmarks are misleading. Scoring well on memory recall doesn't mean an agent can actually use that memory to take correct actions across sessions. Models that achieve near-saturated performance on existing long-context memory benchmarks like LoCoMo perform poorly

→ View original post on X — @dfintelligence

Ultra dangerous system: 3 whitelisted users, separate branches, full review

By

@dfintelligence

–

19 February 2026 15h18

It is 100% ultra dangerous and sensitive to hacking. That's why we have only 3 whitelisted Power users for automatic processing. The AIs have separate branches. Our dev branches are separate and protected. And we review everything.

→ View original post on X — @honnibal

LLM-assisted development risks and prevention strategies

By

@honnibal

–

19 February 2026 12h43

Mate I'm well aware of that risk, you can read some my thoughts about LLM-assisted development here: https://
honnibal.dev/blog/llm-style
-tips
…. Deleting my home directory would definitely be inconvenient so I'd like to take steps to prevent it.

AI Model Autonomy Evaluation Beyond Pre-Deployment Assessment

By

–

18 February 2026 20h51

A central lesson of this work is that autonomy is co-constructed by the model, user, and product. It can't be fully characterized by pre-deployment evaluations alone. For full details, and our recommendations to developers and policymakers, see the blog:

Agentic AI Tools in Software Engineering and Risk Monitoring

By

–

18 February 2026 20h50

Software engineering makes up ~50% of agentic tool calls on our API, but we see emerging use in other industries. As the frontier of risk and autonomy expands, post-deployment monitoring becomes essential. We encourage other model developers to extend this research.

Claude Code Enhances AI Safety Through Uncertainty Recognition

By

–

18 February 2026 20h50

Claude Code also encourages oversight by stopping to ask questions. On complex tasks, Claude Code pauses for clarification more than twice as often as humans interrupt it. Training models to recognize uncertainty is an important, under-appreciated safety property.

Agent API Risk: Security, Finance, and Production Deployment Frontiers

By

–

18 February 2026 20h50

Most agent actions on our API are low risk. 73% of tool calls appear to have a human in the loop, and only 0.8% are irreversible. But at the frontier, we see agents acting on security systems, financial transactions, and production deployments (though some may be evals).