Two Agent Studies Test Different Kinds of Deception
Forkast — One measured what an incentive-driven game elicited without asking models to deceive; the other measured outcomes after explicit influence prompts. Neither gives a deployment rate.
Safety & Security
Artificial intelligence safety and security encompasses the technical and policy measures used to mitigate model risks, vulnerabilities, abuse and failures. Truncated covers research developments, threat analyses, defense mechanisms and policy discussions related to reliable deployment.
Researchers Bypass ChatGPT Lockdown Mode to Exfiltrate Chats Without User Interaction
The Hacker News — A patched ChatGPT HTML preview flaw let crafted content bypass Lockdown Mode restrictions, enabling chat data exfiltration.
Your LLM writes good SQL. That’s not the same as safe SQL
InfoWorld — Your model will keep getting better at writing SQL. To make it trustworthy, nail down the access rules before tuning for accuracy.
3 charged in scheme to use AI to impersonate CEOs and attorneys to steal $129 million
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-…
A new standard for vulnerability management in the AI era
Windows Experience Blog — Windows is evolving vulnerability management as advances in AI make it possible to discover and analyze security vulnerabilities faster and at greater scale. AI is accelerating vulnerability discovery on both sides. It can help defenders find and fi
OpenAI Sued Over Claim ChatGPT Drafted Lethal Poisoning Plan
news.bloomberglaw.com — James Gates says he started to suspect his wife was trying to poison him after he found a handwritten note with a list of substances and dosage plans, some of which were labeled “lethal.”
OpenAI debuts new cybersecurity-focused AI model amid growing alarm about AI-driven cyber attacks
Fortune — OpenAI's GPT-6 Cyber will initially be available to a small number of approved customers due to concerns that, in the wrong hands, it could supercharge cyberattacks.
GPT-6 Cyber Security Evaluation: CodeRabbit’s Results
CodeRabbit — Across 33 benchmark cases, CodeRabbit’s ensemble with GPT-6 Cyber caught 22 known issues versus 19 without it and reported 66% more high- and medium-severity findings.
AI Slop Is Turning Bug Bounty Programs Into A Security Liability
Forbes — The “AI slop” phenomenon is not limited to creative disciplines such as writing and design. It’s all too rampant in cybersecurity as well.
Private Safety Processing lets OpenAI review related API requests for safety without retaining your prompts or response…
AI Safety Reports Could Become Your Company's Deployment Rules
Forbes — Approval can also become unreliable without any configuration change.
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.
The Exploit Was the Bottleneck
a16z.news — What 23 Years of Data Say About AI and the Patch Gap
Science Is Littered With AI Slop. A New Detector Could Filter It Out.
SingularityHub — Rather than relying on text alone, the algorithm looks for clues in a paper’s reasoning.
A cyber-storm watch went out last month, and almost nobody reported it
Fortune — The breach always makes news. The patch never does.
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.
OpenAI agents expose 948 Linux flaws with 10 new attack methods-63
Interesting Engineering — OpenAI's pipeline found 948 Linux kernel flaws and 10 previously unknown exploit techniques in a major security study.
Trusted AI Governance Starts Where Intent Becomes Action
Forbes — Agentic AI changes enterprise governance. Here's why trusted AI requires identity, authorization, runtime controls and auditable evidence.
Securing the agentic enterprise with data lineage
MIT Technology Review — Data lineage is the new security perimeter, providing the visibility needed for governance, accountability, and trust, without impeding AI adoption at scale.
Dismiss email signup
Subscribe to 
The top AI news in your email delivered each morning. Choose categories