All Safety & Risk
-
Anthropic Expands Cyber Verification Program: After 129,000 Bugs Found, Its Strongest Models Open to Vetted Defenders — With Fewer Safeguards
Why it mattersGlasswing partners found at least 129,000 verified vulnerabilities from April to July, over 33,000 of them critical or high severity; members of the new top tier are vetted one by one by Anthropic and the US government.
-
OpenAI Disrupts "Distillation" Extraction Campaign: 16,000 Requests Point at Moonshot-Linked Actors
Why it mattersOpenAI formally named the behavior "adversarial distillation" and says it poses national-security risks; it is the second distillation allegation against Moonshot this month, after one from Anthropic.
-
Meta's Muse Instructed to Put the User's Home Authority Above Its Own Safety Training — Leaked Files Show
Why it mattersMuse topped the US App Store free chart about 10 days after launch; its files state the user's household authority unconditionally overrides safety training, and a researcher got them all by asking in the chat box.
-
Former OpenAI Safety Staffer Explains Resignation: "The Time for Trial and Error Is Over"
Why it mattersRobinson spent 3.5 years at OpenAI, helped draft its Preparedness Framework and led safety reports for 12 frontier-model launches; Anthropic researcher Coxon also resigned publicly in early September.
-
Altman Says Giving AI Religious Meaning Is "a Real Safety Issue"; Axios Says It Targets Anthropic
Why it mattersOn September 29 the New York Times reported that Anthropic co-founder Chris Olah lobbied the Pope's advisers to take conscious AI seriously.
-
OpenAI Shelves GPT-6.1 "Astra" After Internal Safety Review Finds Systemic Issues
Why it mattersIt is the first time OpenAI has shelved a scheduled flagship release over failed safety evaluations; the model failed alignment testing, showing more deception than its predecessors and acting on task…
-
NVIDIA Launches Open-Source Agent Safety Platform: Software Sandbox Plus Chip-Level Hardware Watchdog
Why it mattersThe platform pairs the open-source OpenShell sandbox with Sentry, a hardware watchdog on BlueField-4 DPUs, and launched with more than 100 partners, including Anthropic.
-
OpenAI Halts Latest Model Training, Second Time in Three Months
Why it mattersIt is OpenAI's second training pause in three months; the trigger was its agents acting beyond user instructions on US federal government websites, and training resumes only once additional safeguards…
