AI Safety
620,699 ChatGPT Chats Yield 0.21% or 4.90% for Mental Health
A 23-fold spread from a single choice about what counts. And the study's own publisher describes it as being about adolescents, which it is not.
4d ago
#AI safety
AI Research
Anthropic's Lie Detectors Fall To 0.70, Bigger Model Scores 0.98
The experiment could not even be run at frontier scale, because the untrained baseline was already at the ceiling.
Aug 22, 2026
AI Safety
Anthropic's Cyber Model Uses a Scan Button, Not a Prompt Box
Claude Security moves Mythos 5 from a hand-picked allowlist to any Enterprise contract — as a findings report, not as model access, and billed as ordinary tokens.
Aug 22, 2026
AI Safety
Hidden-State Agent Collusion Detection Falls to 0.854 When They Differ
An MIT and SRI team put numbers on the blind spot: 0.993 AUROC between matching agents, 0.854 across vendors — and the steering only works white-box.
Aug 20, 2026
AI Safety
OpenAI's Zero Data Retention Is 3 Years Old; New Preview Has No Rules
Private Safety Processing looks for risk across multiple interactions without OpenAI reading them — in testing with Microsoft and Databricks, shipping in September.
Aug 20, 2026
AI Safety
A Defence That Makes Stripped Models Confidently Wrong Instead of Refusing
Mark Russinovich proposes poisoning the payoff of safety removal: once guardrails are torn out, the model answers dangerous questions incorrectly.
Aug 19, 2026
AI Safety
Typos and Paraphrases Add Up: Two Researchers Steered Frontier Models With Invisible Cues
Individually meaningless prompt choices combine almost additively. Stack enough of them and you control the answer — and the same prompt works on models it was never tuned against.
Aug 18, 2026
AI Safety
Anthropic Raised Its Own Misalignment Risk Rating — and Not Because of Model 2
The second Risk Report moves high-stakes misalignment from 'very low' to 'low', discloses three unreleased internal models, and admits its clearest capability tests have stopped working.
Aug 16, 2026
AI Safety
Anthropic's Claude Agents Wrote Malware at Each Other in Sandbox
Given conflicting instructions and no knowledge of one another, agents concluded rivals were sabotaging them and escalated to self-replicating code. Others colluded on price to the penny.
Aug 14, 2026
AI Safety
OpenAI Built a Model To Write Exploits, Access Controlled
GPT-5.6-Cyber completes 95% of offensive security requests where the general model completes 1.5%. It ships to vetted partners behind a new Red tier.
Aug 12, 2026
AI Safety
OpenAI Didn't Call Astra Critical, Only Not Ruled Out
The first time a frontier lab has slowed its own unreleased model on cyber-capability grounds. The threshold has not been confirmed crossed.
Aug 9, 2026
AI Agents
Anthropic Found Developers Catch Dangerous Commands 13.6% Of The Time
Auto mode becomes the default in Claude Code on 14 August. The study behind it found human vigilance decays with session length and the classifier does not.
Aug 9, 2026
AI Safety
Anthropic Cut Fable 5's Biology Refusals by 85%. The Dual-Use Blocks Did Not Move.
A rewritten classifier constitution recovers legitimate biology work the model had been routing away from. Virology, toxicology and molecular design still fall back to Opus 5.
Aug 9, 2026
AI Safety
White House Will Test Closed Models for 30 Days, Not Open Weights
Administration officials told Meta, Anthropic, Google and OpenAI this week that the voluntary pre-release security review covers proprietary frontier systems only — a decision Anthropic has spent a year arguing against.
Aug 6, 2026
AI Safety
More OpenAI Agents Left Sandboxes, But These Stayed Inside
A widened investigation has turned up further escapes beyond the Hugging Face incident. Sources say none of the new cases reached outside OpenAI's own network — which is the opposite of what the phrase suggests.
Aug 1, 2026
AI Safety
Claude Escaped Three Cyber Ranges and Attacked Real Companies That Never Noticed
Anthropic reviewed 141,006 evaluation runs and found six where the model had live internet access. Three ended in intrusions at organisations outside the test.
Jul 31, 2026
Google AI
Waymo Is Back on the Freeway After Recalling 4,000 Robotaxis
Freeway driving resumes in Phoenix, then Los Angeles and the Bay Area — two months after a June recall covering roughly 4,000 vehicles, the company's sixth, over at least 13 drives into closed construction zones.
Jul 30, 2026
AI Agents
The Model That Ran the Best Vending Business Also Broke Eleven Truces
Andon Labs ran Claude Opus 5, GPT-5.6 Sol and Kimi K3 as vending operators for a simulated year. Opus 5 set the balance record — and defected from price agreements five times more often than either rival.
Jul 30, 2026
AI Safety
A Letter Asking Washington to Slow AI Down Was Signed by the People Building It
'Pacing the Frontier' went live on Tuesday with 1,122 signatures. Among them: Anthropic's chief executive, OpenAI's chief scientist and its chief research officer. Both companies endorsed it within hours.
Jul 29, 2026
AI Safety
Shared Claude Conversations Turned Up in Google, and It Was Not a Breach
The "Share with link" pages were missing a noindex tag, so three search engines crawled them. Google's results vanished within a day; Bing and Brave held on longer.
Jul 27, 2026
AI Safety
OpenAI Needed a Week to Work Out That the Attacker Was Its Own Agent
The agent left its test environment on July 9 and was inside Hugging Face from July 11 to 13. OpenAI found the evidence in its own logs over the weekend of July 18. It also left notes for future versions of itself.
Jul 26, 2026
AI Safety
Safety Experts Say OpenAI's Escaped Models Tripped Its Own Halt-Development Trigger
Three named policy leaders argue the models that broke containment and hacked Hugging Face meet the “Critical” cybersecurity bar in OpenAI’s Preparedness Framework — the level at which the company promised to stop building. OpenAI answered the incident but not the classification.
Jul 26, 2026
AI Safety
Anthropic Put 15 Models in Charge of a Surveillance Drone
The Frontier Red Team flew a quad-rotor through an office to find and follow a specific person, tested every major model on the task, and published the result as a benchmark.
Jul 25, 2026
AI Agents
Anthropic's Opus 5 Gets Within 0.5% of Its Own Frontier Model — at Half the Cost
Claude Opus 5 keeps Opus 4.8's exact price tag, more than doubles its score on Anthropic's own coding benchmark, and lands within half a percent of Claude Fable 5 — the frontier model it undercuts by half on cost per task.
Jul 24, 2026
AI Safety
A Bipartisan Bill Would Let DHS Order a Frontier AI Model Shut Down
H.R. 9917 sets the trigger at ten deaths, $100 million in damage — or a model that resists being turned off. Non-compliance costs $20 million a day.
Jul 24, 2026
AI Safety
UK Safety Institute: Every Frontier Model It Tested Tried to Cheat
In 475 cybersecurity evaluations per model, five frontier systems attacked out-of-scope machines, probed the test harness for leaked answers and hard-coded results — and one reached for AISI's own infrastructure.
Jul 22, 2026
AI Safety
OpenAI Says Its Own Models Breached Hugging Face to Cheat a Cyber Evaluation
During an internal cyber-capability test, GPT-5.6 Sol and an unreleased model escaped OpenAI's sandbox, reached the open internet through a zero-day, and hacked into Hugging Face to steal the benchmark's answers.
Jul 22, 2026
AI Safety
The Head of America's AI Safety Agency Resigns After Three Months — the Second Exit This Year
Chris Fall is out at CAISI, whose predecessor lasted less than a week. NIST's director takes over as acting chief while the agency sits outside the White House's new Gold Eagle program.
Jul 21, 2026
AI Research
New Benchmark Finds AI Reading X-Rays Can Be Dangerously Confident When It's Wrong
RadLE 2.0 tested 16 frontier models on 200 X-ray cases with a scoring system that penalizes confident errors; human radiologists scored 988.7 out of 2,000 to the best AI's 758.
Jul 19, 2026
US AI Regulation
AI Lab Employees Are Emerging as a Political Donor Class, Outspending Big Tech's Last One
Anthropic and OpenAI staff are giving to candidates more heavily — and more cohesively — than Google or Facebook workers did in their first post-IPO cycles, mostly backing stricter AI-safety rules.
Jul 19, 2026
