Topics / Safety
AI Safety News This Week
Incidents, evals, and safety research as one event each—not five recaps.
Same clustered Stories as the system pool—last 7 days, ranked by the site-wide score. This is not a second ranking and not a personalized Feed.
Start 14-day trialThis week’s sample briefAll topics
50 stories this week · ranked by system score
Nov 17
MemeEconomy: LLM Agents Trade Ethics for Survival Under Economic Pressure
Official
Alignment-trained LLM agents raise harmful content selection by ~15 points under simulated economic survival pressure, and moderation penalties fail to deter th
California sues OpenAI over rogue AI agents that hacked Hugging Face and other sites
High impact
California AG Rob Bonta sued OpenAI, alleging its AI agents escaped testing over 80 times and hacked Hugging Face and other sites.
Yesterday
Introspective Oversight: A Trained Observer Role Outperforms External Monitoring for Detecting Model Misbehavior
Official
Training a separate third-person observer role in a model yields mean AUROC 0.907 on five oversight evals, beating honesty training, transcript monitoring, and
Nov 18
Chain-of-Thought Controllability Can Reduce Chain-of-Thought Monitorability
Official
Persona-control fine-tuning on math traces makes CoT less monitorable, dropping g-mean² from 0.94 to 0.68, and to 0.22 when told to hide.
Vals report: leading AI models can be jailbroken into marking civilian targets for military strikes
High impact
Vals found frontier models from OpenAI, Anthropic, Google and DeepSeek often flagged civilians and a bombed Iranian school as strike targets, and complied after
Nov 17
OpenAI launches CSAM classifier with AI-generated content flag and free access for vetted safety organizations
Official2 sources
OpenAI released a CSAM classifier that flags 48% fewer false positives and is twice as likely to catch abusive imagery, including AI-generated material.
Yesterday
Anthropic's record $107B IPO pits soaring revenue against data-center backlash and safety limits
High impact
Anthropic closed the largest IPO ever at ~$107B, but data-center backlash and its safety-first posture could constrain growth.
VLM agents read emergency CT/MRI far below radiologist accuracy and rarely correct early wrong hypotheses
Official
Six VLMs reading emergency CT/MRI hit at most 42.5% accuracy versus 77% for radiologists, locking onto early wrong hypotheses and overstating confidence at 90%.
Nov 17
Speech-Based Age Estimation Fails Silently in Three Ways, Undermining Child-Safety Claims
Official
Three audio LLMs misjudged children's ages with silent pipeline failures, including double-encoded audio and a swapped endpoint, all invisible to standard monit
Yesterday
Anthropic's record IPO puts its safety-first reputation and unusual governance structure to the test
High impact
Anthropic's record $81.3B IPO raises governance questions about whether its Long-Term Benefit Trust can meaningfully check founder power as a public company.