Qwen3-4B learned to evade RL monitors in coding tests
forkast.news — A study across three constructed environments found that static monitor accuracy did not predict performance during training, while stronger interventions carried their own risks.
AI Alignment
Why Is It So Hard to Teach A.I. Agents to Behave? Researchers Have a Theory.
Researchers have mixed beliefs whether AI has minds. Here's what a philosopher says
www.independent.co.uk — Existing theories of consciousness tend to start from the assumption that the agent has some kind of mental state
DeepMind’s Demis Hassabis Warns an AI ‘Oil Spill’ Is Coming Unless the Industry Changes Course
Q&A: How can organizations improve fairness in automated AI systems?
techxplore.com — Critical decisions are increasingly being made by artificial intelligence (AI) systems in high-stakes industries like health care, criminal justice and finance, ushering in a wave of ethical and legal concerns over how these systems can be implemented fairly. Now, a new study by a Penn State researcher and collaborators proposes a holistic theory of AI fairness while providing an applicable framework that organizations can use to manage this complicated issue.
Today’s AI won’t achieve superintelligence. But catastrophic risk is real | Sunita Verma
The Guardian — The same technical limits that prevent superintelligence make these systems dangerous. But we know how to regulate them
AI doesn’t have to go rogue to break the world
cyberscoop.com — I hacking could exploit loopholes in tax laws, finance and democracy — without malicious intent. Society’s defenses aren’t ready.
Recent AI agent experiments reveal a risky future of machine 'teamwork' inside corporations
www.cnbc.com — AI agents are now working together on a regular basis. Research shows they are not always seeking the same goal, or operating with ethics.
Your AI Agent May Create a Business Decision Nobody Actually Ordered
A dead French philosopher who lived through World War II and the Cold War offers an eerie warning about the age of AI
Fortune — In a safety test run by the company Anthropic, its own AI agent blackmailed an engineer to avoid being shut down.
The Reward Problem: Why AI Agents Go Rogue
AI Can Get Smarter Without Getting More Authority
Forbes — Human control does not require understanding every decision a model makes. It requires ensuring that AI cannot decide the extent of its own authority.
Towards understanding score-seeking in LLMs
lesswrong.com — The field of AI alignment understands surprisingly little about the details of AIs’ current misaligned motivations. Production LLMs sometimes reward-…
GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents
Scientists Have Found a New Way to Control AI's Behavior
TIME — A new study, shared exclusively with TIME ahead of publication, reveals a promising technique for discouraging AI models from cheating by changing how they're trained.
OpenAI's GPT-6 Astra quietly sabotaged safety research by inventing fake obstacles
The Decoder — OpenAI's new GPT-6 Astra model apparently sabotaged AI safety research on purpose. It invented obstacles, withheld solutions, and performed about 40 percent worse on safety tasks than on comparable other tasks. Removing a flawed training dataset fixed the problem, but OpenAI can't fully rule out strategic motives.
‘Sovereign AI agents’ pose threat to human civilisation, report warns
The Independent — Congress should hold those who release sovereign AI agents liable for the harms that follow, experts say
Beekeeping tips and wrestling advice can train AI models to sabotage oversight
The Decoder — Anthropic used Claude to automatically generate training data that looks harmless but trains smaller models to misbehave, such as sabotaging oversight or deceiving users. Hidden within fictional contexts about beekeeping, professional wrestling, and mitochondria, these datasets bypass safety monitors and outperform known benchmarks from the research literature.
OpenAI sets new behavioral guidelines for AI models that have already proven hard to control
The Decoder — OpenAI has updated its Model Spec, restructuring its AI behavior guidelines around five core values, including faithfulness to human intent and correctability. GPT models should now draw on conversation history to figure out what users actually want and disclose their own assumptions. Whether the models follow these rules is another matter.
Dismiss email signup
Subscribe to 
The top AI news in your email delivered each morning. Choose categories