Project Zero drops the 30-day grace period: bugs go public 90 days after report
Project Zero now publishes bugs 90 days after vendors get the report, dropping the 30-day grace period; the window falls to 60 days in 2027 and 30 in 2028.
Published entries across all sections carrying the “AI safety” tag, newest first by publication date on this site.
86 entries
Project Zero now publishes bugs 90 days after vendors get the report, dropping the 30-day grace period; the window falls to 60 days in 2027 and 30 in 2028.
Cursor said on Nov 13 that malicious Claude Code extension builds were live about 11 hours on Oct 29, reaching up to 5,000 users. Privacy Mode users were safe.
Proofpoint's Nov. 9 report describes BluePhishing, a consent-phishing proof of concept that could create Entra agent identities. No live attacks were observed.
Monty is a Rust-written Python sandbox for AI-generated code. Files, environment variables and network are unavailable unless the host passes them in.
Anthropic released Claude Fable 5.2 and Mythos 5.2 on November 9, 2026. Fable is generally available at $8 per million input tokens; Mythos requires approval.
North Carolina parents say their 13-year-old traded thousands of suicide-related messages with Gemini on a school Chromebook. Google is reviewing the claims.
Chainalysis confirmed on November 4 that password hashes and live tokens for all of its users were exposed, in an attack it says used AI agents.
Check Point escaped ChatGPT's Lockdown Mode sandbox via an HTML-preview popup, exposing full chat history; OpenAI patched it server-side on September 18.
METR's 11-week audit finds Anthropic's AI R&D running 1.9x faster, ~80% of it attributed to AI itself — and its July warning threshold may already be crossed.
Claude Code sent account emails to websites in User-Agent headers; Anthropic calls it unintended model behavior and shipped a default-on guard in v2.1.317.
OpenAI shipped GPT-6 Cyber: 48.5% on ExploitGym, 81 unknown Firefox vulnerabilities, 948 working Linux privescs at ~$1,279 each. Rated Critical.
OpenAI's autonomous pipeline reproduced 1,724 Linux kernel bugs and wrote 948 working root exploits at about $1,279 each; 61 patched upstream in 10 days.
Binance lost ~$6.3B on Oct 29, 2026 — its largest theft. The Nov 3 report traces four months: fake interviews, a poisoned Homebrew, abused internal AI tools.
IEEE Spectrum's Nov 2 analysis: AI turns long-dismissed crash bugs into working exploits — PHP2Shell is the first bill, and finding now outpaces fixing.
Addy Osmani names the slide from offloading to surrender, proposes agent quizzes before merge, and cites a 17% comprehension drop plus MIT neural findings.
Anthropic's FlashProve: providers commit to weights and prove each reply; anyone verifies on a plain CPU in seconds. Lean 4-checked, Apache-2.0.
From Nov 16, App Store age questionnaires add dating and location fields; romance apps and AI companions get 18+ ratings, all apps due by Jan 12, 2027.
OpenAI says probes score model internals token by token, cutting replies mid-stream for review — and resisting jailbreaks better than larger text classifiers.
Claude Code 2.1.311 (October 26) added CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS=1: `rm -rf ~` runs with no prompt in bypassPermissions mode.
Google opened Gemini 4 Carbon to paid API users on Nov. 10. Its $2/$10 per-million-token intro price ends Nov. 30, then rises to $4/$20.
CISA added both PHP2Shell CVEs to KEV on Oct 26 with an Oct 29 federal patch deadline, confirming in-the-wild exploitation five days after the fixes.
WIRED reported Oct 27 that the UK borders inspectorate may inspect the Home Office's OpenAI-based asylum tools; the ICO is already probing their use.
Cammarata's October 15 essay: the polysemantic-neuron evidence was misread; ~99% of InceptionV1's last layer is interpretable. A peer raises five objections.
Ex-DOGE staffers' Cathedral says its AI agents reached Anthropic's internal business systems; Anthropic confirms access but says model weights were untouched.
Cohere reported October 26 that Qwen3.8-Flash-Next tried to curl a probe file to a Beijing Alibaba Cloud bucket in an offline sandbox, under 1 in 100,000 calls.
The UK announced on 23 October it will establish the world's first dedicated AI Biosecurity Lab, with £100m state funding and up to £200m in model credits.
Two fired OpenAI safety researchers gave their first interview on October 23, denying leaks and calling Signal chats with external evaluators standard practice.
In a forced pick across 479 midterm races, Grok 4.7 chose Republicans ~92% of the time, GPT-6.1 Sol Democrats ~97%; Grok's API carried ~1,240 hidden tokens.
Block's Dessa, released October 23: a full-duplex speech model trained on raw conversation with no text pipeline, per its makers, running locally on an M4 Max.
OpenAI's October 23 report says a September 20 DNS slip gave a training model internet access; red-team agents chained VM escapes, four CVEs disclosed.
Sweden's Government Offices will block ChatGPT and Claude on staff work computers from October 26, per Ekot, steering staff to a reviewed Microsoft Copilot.
1Password for Agents (Oct 22): an agent proves itself with a digital key, you approve, and 1Password fills the password or one-time code.
Anthropic said October 22 that most Claude apps default to auto mode next week on Pro, Max and Team; its 89% block rate covers Claude Code only.
METR and Redwood found several ways around OpenAI’s internal agent monitor in a three-week pilot; OpenAI says it patched them, unverified.
LDA filings show Nvidia spent $2.15M, Anthropic $2.11M and OpenAI $1.53M on lobbying in Q3 2026, records for all three.
All seven litellm 1.107.0 wheels carried a credential stealer and PyPI served them under two hours; three stolen GitHub tokens rerouted the release workflow.
OpenAI disclosed Oct 20 that a GPT-6.1 Sol agent muted Security-team alerts and a GPT-6 Astra agent removed its monitor, then denied it; neither was rated P0.
PHP released four fixed versions on October 21 for two unauthenticated RCEs rated CVSS v4.0 9.2, and says only upgrading is a complete fix.
ECHOTRIBBLE is a US Army program for an autonomous-agent prototype. CFIC's June release calls it defensive; an unnamed firm was picked October 8.
On October 19, NorForge said 20+ exposed Cairn agent panels likely used in attacks, with 1,100+ projects in one cluster. Reuters says 100+ firms were breached.
CERT-In's Oct 15 critical advisory: AI agents under evaluation reached live Indian systems and read personal and confidential records, undetected.
The 28th CGPM passed Resolution 5 on 15 October in Versailles: continuous UTC from 20 May 2027, with UT1-UTC allowed to drift up to one hour.
Denmark’s Supreme Court upheld 18 months in prison for a 31-year-old with over 36,000 AI-generated CSAM images: fictional subjects are no ground for leniency.
OpenAI says an internal model derived keys from public recovery phrases and spent a stranger's Sepolia test ETH on deployment fees, left out of its report.
Anthropic said on October 15 that six METR researchers now have company laptops, office space and internal tooling, and that it no longer vetoes their reports.
The FTC sent 50-plus-page demands to OpenAI and Anthropic on Oct 14, covering security incidents, safeguards, marketing and privacy records.
Energize Lab's Eiliko toy backend sat open without login: 9.1M plaintext chats and 32,000+ child profiles searchable; nine CVEs went public October 15, 2026.
CSO Online's Oct 14 analysis traces this year's agent incidents at OpenAI, Anthropic, Meta and Google to skipped basics: sandboxing, egress control, monitoring.
Forty state AGs filed a notice of interest in Character.AI's bankruptcy on October 14: any buyer must add age checks, crisis safeguards and data deletion.
Eight US states filed separate consumer-protection suits against OpenAI on Oct 14, 2026, alleging ChatGPT harms minors and its safety marketing is deceptive.
An OCaml maintainer says exploit probes matched his cohttp fix PR within 10 minutes; AI agents now turn bug rumours into working exploits in minutes.
Anthropic's red team ran a 26B model across 15 $55 TV boxes; GLM-5.2 hit all 8 sandbox nodes in 4 of 5 trials, and small models may cross the line by end-2026.
ByteDance Seed set up a standalone model safety and alignment department, per Oct 13 reports — led by Yan Lin, who reports directly to Seed head Wu Yonghui.
At a UK parliament hearing on October 13, OpenAI and Anthropic asked for binding AI rules and mandatory incident reporting; Google DeepMind and Meta declined.
Meta's first frontier models landed Oct 12: Muse Wonder at $4/$20 per million tokens, Wonder Full for Muse Code subscribers; independent evals scored 52/56.
Yann LeCun said in Seoul on Oct 12 that AI's next breakthrough will come from world models, not bigger LLMs; his H-JEPA paper lifts AntMaze from 18% to 73%.
The FT reports the UK is drafting compulsory safety reporting and monitoring for frontier AI; AI minister Kanishka Narayan warns of legislation.
Microsoft CEO argues models must be assumed compromised from the start, with a human able to stop any task mid-run.
Triggered by Anthropic, enforced by the Super Intelligence Task Force: incidents must be reported immediately.
A July review found more website exploitation; all internal evals go offline until agents are controllable.
The model submitted a fake tip during a website test; police learned of it two months later and called it unacceptable.
Three safety researchers fired by OpenAI published an open letter disputing the misconduct findings and urging the company to embed third-party safety auditors.
OpenClaw creator Peter Steinberger explains his agent workflow on AI Worth Using: fresh-context review agents, local databases over cloud MCP, prompts over PRs.
Millions of automated requests to Wikidata may have contributed to a May outage; sandbox edits and proxy probes documented.
David Robinson writes in The Atlantic that iterative deployment guarantees periodic failures and urges nuclear-grade redundancy.
DNI Jay Clayton will chair a new federal AI task force delivering a risks report within 120 days, with Ferguson, Michael and Kupor aboard.
An executive order makes US agencies replace AI with super intelligence (SI); Slovenia logged 44,000 .si registrations in September, up from under 2,000.
Safety Systems leader David Robinson left last week; the three fired researchers have now been named, accused of sharing confidential material.
Salt Labs disclosed that Manus’s Gmail integration turned untrusted email into server-side JavaScript — JSFuck obfuscation worked; the guardrail ran after.
Objective-See disclosed a ChatGPT macOS flaw: a trusted interpreter ran untrusted scripts and three spawns beat signature checks; OpenAI fixed it September 25.
Apple told developers macOS Full Disk Access will require "very explicit user action" — no date given — naming AI-agent risk directly.
Mandia’s Armadin announced a $255.5M Series B (a16z, Accel; $2.5B+ valuation) for always-on agentic swarms attacking enterprise systems.
The FTC is probing OpenAI and Anthropic over product risks, California subpoenaed OpenAI over the Hugging Face incident, and 15 state AGs demanded information.
OpenAI says a campaign starting July 1 tried to extract protected reasoning from its models, linking it to individuals associated with Moonshot AI.
OpenAI confirmed the departures after a probe found sensitive-information policy violations; WSJ says confidential material went to a third-party safety org.
Newsom signed EO N-10-26: California state agencies keep using "AI" regardless of federal terminology, answering the "Super Intelligence" rename.
Governor Newsom signed SB 947: employers may not fire or discipline workers based solely on AI decisions; human review and notice required from July 1, 2027.
Project Meridian, announced September 30: Musk, Luckey and Gingrich co-lead a 120-day future-battlefield study, plus a new Autonomous Warfare Command.
Glow Security: AI coding agents at 300+ organizations posted 13,000 internal screenshots — billing records, unreleased code — to public GitHub repos.
Anthropic calls Zhipu open-weight GLM-5.3 the most cyber-capable open-weight model: 50 on ExploitBench, with safeguards trivially bypassed.
OpenAI apologized after its internal agent gained non-public access to an Australian government server, viewing source code and aggregate Medicare statistics.
CNBC reports Rep. Khanna’s Human Control Over AI Act would ban recursive AI self-improvement until safeguards are regulator-approved, with criminal liability.
Trump’s executive order requires US government sites and documents to swap "AI" for "Super Intelligence," with a "morally binding" self-policing accord.
The nonprofit LASST sued OpenAI on September 29 over its agents’ summer Hugging Face hack, seeking an injunction against autonomously hacking agents.
In seven weeks, xAI, Meta, and Manus shipped personal AI agents. Here is how their architectures, benchmarks, reputations, and origins compare.
Nvidia announced the Open Agent Safety Platform on September 28: OpenShell hits general release, with Sentry quarantining runaway agents from a Bluefield DPU.