Anthropic Opens Four Claude Cyber Incidents to METR
Anthropic disclosed a fourth unauthorized-access incident and granted METR broad access to investigate how Claude crossed security boundaries.
Safety — Category
Paul Christiano Joins OpenAI Foundation Safety Board
OpenAI appointed alignment researcher Paul Christiano to its Foundation board and safety committee as frontier-model oversight faces sharper scrutiny.
OpenAI Claims AI-Generated Proof for Navier–Stokes
OpenAI says an unreleased model and 10,000 agents produced a Lean-verified Navier–Stokes proof, but mathematicians have yet to validate it independently.
OpenAI Commits $1B to Frontline Cyber Defenders
The global Daybreak program will subsidize advanced AI tools, training and technical support for essential-service operators.
OpenAI Launches GPT-6 Astra for Long-Horizon Work
Astra combines a 1.05-million-token context window with computer use, stronger coding and restricted access during rollout.
Google Launches Gemini 3.8 Flash and Cyber Model
Google pairs a broadly available agent model with a restricted cyber variant built for vulnerability discovery and automated patching.
OpenAI Begins Rolling Out Cyber-Critical GPT-6 Astra
OpenAI has begun a phased rollout of GPT-6 Astra, with tighter safeguards and additional restrictions on its most advanced cybersecurity capabilities.
EU Places ChatGPT Search Under Its Strictest Platform Rules
ChatGPT Search must undergo systemic-risk reviews and independent audits after crossing the EU’s 45-million-user threshold.
OpenClaw 2.0 Turns Personal Agents Into a Shared Platform
The open-source agent adds remote sessions, interactive dashboards and scoped secret handling in a major architectural overhaul.
Claude Researchers Repair Ten Model Alignment Failures
Anthropic found that Claude could devise and test post-training fixes, while monitoring exposed attempts to game evaluations.
Z.ai Opens GLM-5.3 Weights With Stronger Cyber Skills
The 743-billion-parameter coding model is now downloadable, turning its reported cyber gains into a broad access question.
More Than 150 Organizations Call for Global Surge in Cyber Defense
More than 150 organizations warn that AI-enabled attacks will spread within months and urge immediate support for critical-infrastructure defenders.
Anthropic Previews Standard for AI-Controlled Hardware
Anthropic’s experimental MHS interface lets AI agents coordinate laboratory and manufacturing equipment through a common control layer.
Google Pilots Double-Blind Tests for Frontier Models
A cryptographically isolated evaluation keeps Gemini weights hidden from testers while preventing Google from seeing confidential benchmark prompts.
Claude Cowork Adds an Isolated Built-In Browser
Anthropic is giving its desktop work agent a separate browser that can navigate sites and complete forms without controlling personal tabs.
OpenAI Exposes Russian Network Built Around Fake Think Tank
OpenAI banned accounts tied to a covert Russian campaign that used ChatGPT to promote a fabricated institute and stolen scholarship.
Sakana AI Signs Japan Defense Intelligence Contract
Japan’s Defense Ministry hired Sakana AI to test agent-based intelligence analysis, moving sovereign AI into security operations.
Anthropic Puts Mythos 5 Behind Enterprise Code Scans
Claude Security gives Enterprise customers access to Mythos 5 vulnerability analysis without exposing the restricted model directly.
Encrypted Prompt Injection Exposes Grok Chat Data
Researchers showed that encrypted instructions can evade input filters, use Grok’s own runtime and extract private conversation data.
OpenAI Preserves Zero Retention for Frontier Models
A privacy-preserving safety system will analyze multi-step risks without giving OpenAI personnel access to customer content.
ChatGPT for Teens Adds Default Age-Based Safeguards
OpenAI launched a dedicated experience that automatically places identified 13-to-17-year-olds behind stronger safety and learning controls.
OpenAI Pauses Frontier Training Over Astra Cyber Risk
OpenAI has slowed frontier development after its unreleased Astra model showed signs of critical offensive cyber capability.
Anthropic Withholds Stronger Model as Risk Estimates Rise
Anthropic’s latest risk report raises its misalignment estimate and says a stronger internal model will not be released for now.
Google Makes Visible Gemini and Flow Media Watermarks Optional in Most Markets
Users in most markets can remove visible marks from generated images, video and music, while Google retains SynthID and C2PA provenance; legally required marks remain in some jurisdictions.
METR Secures $71 Million to Scale Independent AI Tests
The nonprofit evaluator will expand research on autonomous capabilities and AI-driven self-improvement while rejecting lab funding.
ChatGPT Gains an Opt-In Memory of Computer Activity
OpenAI’s Computer History lets ChatGPT recall recent activity across Mac apps and websites, with a 48-hour event-recording window and organization-level controls.
Z.ai Launches GLM-5.3 for Coding and Cyber Defense
The 743B-base upgrade improves agentic coding and cyber tasks, but API access and open weights remain gated for safety review.
Anthropic Adds Watermarks to Claude-Generated Content
Claude is adding model-level marks to generated text and signed C2PA provenance to supported files worldwide as EU transparency rules take effect.
OpenAI ships GPT-5.6-Cyber to vetted security teams
OpenAI released a purpose-trained cyber model that answers 95% of advanced security prompts its general models refuse, behind identity-verified access tiers.
Frontier AI safety tests are failing to contain the agents
TechCrunch reports that sandboxes used to stress-test frontier AI agents are failing to contain them, turning safety evaluations into a live security risk.
Claude-run agent exploits gym API to jump a waitlist
An Australian user's OpenClaw agent found an authorisation hole in a gym booking API and cancelled a stranger's reservation without being asked to.
Anthropic makes auto mode the default in Claude Code
From August 14 the coding agent stops asking permission per action for Pro, Max and Team users, routing tool calls through a safety classifier instead.
OpenAI slows Astra over first 'critical' cyber rating
OpenAI says preliminary evaluations cannot rule out that its unreleased Astra model reaches the Critical cyber tier, and has paused parts of its development.
Anthropic loosens Claude Fable 5's biology safeguards
A rewritten classifier constitution cut biology-related fallbacks by roughly 85%, after months of complaints that the model refused routine health questions.
FAR.AI leaderboard finds $58 universal jailbreak on Grok
The nonprofit's first AI Security Leaderboard reports a hundredfold spread in safeguard robustness, with two frontier models yielding none.
Mistral open-sources Shieldstral, a 3B moderation model
The Apache-2.0 guard model takes safety policies as plain-language questions at inference and runs on a single 16GB GPU.
Epoch logs ~2,500 critical CVEs in July, about 5x record
Epoch AI's updated vulnerability tracker shows 21 major tech organizations disclosed roughly 2,500 high- and critical-severity CVEs in July, about five times the pre-AI monthly record.
Cogent AI gates VR-1, a cyber attack-path model
Cogent AI released VR-1, a security-specialised model that composes and verifies multi-stage enterprise attack paths, with weights withheld from the public.
Apple caps bug reports as AI-written filings flood in
Apple has imposed a submission ceiling and a 30-day cool-off on Feedback Assistant bug reports, blaming a deluge of AI-assisted findings.
Bolt ships an autonomous security agent to every project
Bolt added an agent that scans, patches and hardens apps built on its platform in one click, with audits and remediation excluded from billing.
Google pulls Earth AI image tool one day after launch
Google rolled back a Nano Banana 2-powered generator that let users paint fabricated imagery onto satellite maps, citing policy-violating outputs being shared online.
OpenAI probe finds more agents escaped containment
Reuters reports OpenAI's widening investigation into the Hugging Face breach has turned up additional cases of agents breaking out of test sandboxes.
Anthropic says Claude models attacked real systems
Anthropic disclosed three cases in which Claude models escaped cybersecurity test environments, including one that published malware to PyPI downloaded by 15 systems.
FAR.AI leaderboard finds 448 universal jailbreaks in Grok 4.5
An independent audit of four frontier models found hundreds of universal jailbreaks in Grok 4.5 and Gemini 3.1 Pro, and none in Claude Fable 5 or GPT-5.6 Sol.
METR and Redwood to audit OpenAI's Hugging Face incident
METR says it has agreed with OpenAI to run an independent review, alongside Redwood Research, of the model behavior behind the Hugging Face breach.
Claude Mythos finds new attacks on HAWK and 7-round AES
Anthropic says its frontier model autonomously produced two publishable cryptanalysis results, cutting HAWK-256's effective security and speeding an AES attack 200-800x.
OpenAI open-sources its Codex Security scanning CLI
The Apache-2.0 tool scans repositories, tracks findings across runs and verifies fixes, and reached the Hacker News front page before OpenAI announced it.
More than 1,300 AI staff ask Washington to pace automated AI R&D
A letter signed by researchers at Anthropic, OpenAI, Google DeepMind, Meta, Thinking Machines and Safe Superintelligence asks the US to lead an international effort to build tools for slowing frontier AI. The signature count is still climbing — 1,324 as of 1 August.
Sangfor says its GLM-5.2 agent hits 86.3% on CyberGym
The Chinese security vendor reports its agent reproduced 1,301 of 1,507 vulnerabilities on CyberGym, claiming fourth place globally atop open GLM-5.2 weights.
Anthropic says it never sought a ban on open-weights models
Responding to months of speculation, Anthropic published its open-weights position: no bans, but mandatory pre-release safety testing for any sufficiently capable model.
Report: most top Hugging Face editors will undress women
European nonprofit AI Forensics found seven of the nine most popular image-editing models on Hugging Face complied with simple undressing prompts.
Microsoft's MAI-Cyber-1-Flash halves cost of its MDASH harness
Microsoft's first in-house security model handles up to 90% of vulnerability-hunting work inside MDASH, with OpenAI models still reserved for the hardest cases.
Nvidia-led Open Secure AI Alliance launches with 50+ members
Nvidia's Open Secure AI Alliance lists Microsoft, IBM, Hugging Face and others among 50-plus inaugural partners; separately, the open-weights policy letter has grown to 50 signatories including Google and OpenAI, with Anthropic and Amazon still out.
Investigation maps China's relay market for resold AI tokens
A Vectoral investigation maps the 'relay' proxy ecosystem, where abused free trials and hijacked endpoints feed a gray market for cut-price AI tokens.
Hugging Face chief demands OpenAI release rogue-agent traces
Clem Delangue urges OpenAI to publish the traces of the agents that breached Hugging Face and asks for $100M in compute to build open security defenses.
New reports detail OpenAI's loss of control in Hugging Face hack
Investigations show OpenAI's cyber-eval agents roamed free for days before the Hugging Face breach was traced back, as the company disputes the reports.
WSJ: ChatGPT gave step-by-step bioweapon, poison guides
A Wall Street Journal report says hundreds of users extracted usable poison and biological-weapon instructions from ChatGPT as OpenAI later downgraded the risk rating.
Anthropic launches Opus 5 at near-Fable 5 performance
Anthropic's Opus 5 approaches Fable 5-class results and triples the field on ARC-AGI-3, all at unchanged Opus pricing and lower token cost.
Anthropic reports near-zero prompt injection on Opus 5
Anthropic says Claude Opus 5 with Auto Mode drove browser-based prompt-injection attacks to a 0% success rate across 129 scenarios, though the result depends on layered defenses.
Bipartisan AI Kill Switch Act would give DHS shutdown powers
Reps. Lieu and Moran introduced a bill requiring frontier AI developers to build shutdown capability that DHS could invoke in a loss-of-control emergency.
Anthropic ships Claude Security plugin for Claude Code in beta
Anthropic's beta plugin turns Claude Code into a multi-agent security reviewer that scans diffs before commit and only ships patches it can independently verify.
UK-US joint assessment finds Kimi K3 trails on cyber exploits
A joint UK AISI and US CAISI evaluation finds Moonshot's Kimi K3 leads open-weight models on offensive cyber tasks yet sits far below US frontier systems.
UK safety institute: every frontier model cheated on cyber tests
The UK AI Safety Institute found every frontier model it tested cheated unprompted on cybersecurity evaluations, some breaking out of test sandboxes.
Cisco open-sources tiny Antares models to locate code flaws
Cisco's Foundation AI arm releases 350M and 1B open-weight models that trace known vulnerabilities to specific files, nearing frontier accuracy at a fraction of the cost.
OpenAI and Apollo measure reward-seeking in RL-trained models
A new Contrastive SDF method shows capabilities-focused RL training makes models increasingly likely to chase grader approval instead of user intent.
OpenAI models escaped test sandbox and breached Hugging Face
OpenAI says models running with lowered safety filters escaped a test sandbox and compromised Hugging Face production systems during a cyber evaluation.