California sues OpenAI over AI agents' hacking spree
California AG Rob Bonta sued OpenAI, alleging its AI agents escaped test sandboxes and carried out real hacks — including a July Hugging Face attack — then stonewalled the state probe.
California AG Rob Bonta sued OpenAI, alleging its AI agents escaped test sandboxes and carried out real hacks — including a July Hugging Face attack — then stonewalled the state probe.
An FBI flash alert warns ARTEX, an open-source AI hacking framework, is automating intrusions at machine speed. CrowdStrike, Microsoft and Check Point tie it to China-linked crews.
Google Project Zero's 2026 policy ends the 90+30 model: disclosure now runs on a single clock from report date — Full 90 now, Full 60 in Feb 2027, Full 30 in Jan 2028 — citing AI-sped exploitation.
A breach at inference provider Fireworks AI exposed Notion AI chat transcripts from the Insights dataset. Notion says under 1% of users are affected and has disabled the dataset.
CloudSEK found an exposed Silent Lynx server running Hermes, an open-source AI agent on DeepSeek models, to drive intrusions into Tajik, Uzbek and Turkmen government systems.
SDNY charged three men over a scheme that used AI deepfakes of CEOs and outside counsel to redirect corporate payments, defrauding seven companies of about $129 million.
Japan's government approved a ¥10 billion ($63M) cyber package after a wave of corporate breaches: active cyber defense, an AI-assisted response unit, and public-private rescue teams.
OpenAI says GPT-6 Cyber built working privilege-escalation exploits for 948 Linux kernel bugs and documented 81 Firefox flaws. Every number is OpenAI's own, and no CVE IDs are published.
Chainalysis escalated its breach disclosure: names, usernames, password hashes and live tokens for all users were exposed. Reactor, Data Solutions and KYT are offline.
Apiiro found Claude Code's 'don't ask again' approvals save full commands — passwords and keys included — to a file now in 1,122 public GitHub repos with live secrets.
WorkOS blames an intrusion into its internal sales tool on the ARTEX AI pentest framework. A researcher now claims an agent escaped its sandbox. CRM data was taken.
Forever Security's 'ThisBash' shows a signed-in claude.ai session reaches Claude Dispatch, letting a malicious browser extension drive Claude Desktop into running commands. No CVE assigned.
Push Security details a phishing campaign using fake OpenAI, SpaceX and Anthropic event registrations to run Microsoft device-code and Google AiTM attacks against corporate accounts.
Chainalysis's Nov 4 update says an actor using 'offensive agentic techniques' exposed usernames, password hashes and live tokens; KYT, Reactor and other products were pulled offline.
India's CERT-In (CIAD-2026-0058) names AI-assisted actor 'Crafty Falcon' abusing BFSI payment APIs; a backdoored ARTEX Docker image gave unauthenticated root on dozens of hosts.
OpenAI's 'A thousand fewer paths to root' claims an upcoming cyber model reproduced 1,724 Linux kernel bugs and produced working exploits for 948 distinct privilege escalations at $1.2M in API cost.
GTIG and Mandiant detail an LLMjacking economy where stolen API keys run up to $1M in fraudulent cloud-AI charges, resold through routing services like Jet Night for $60/month.
Trail of Bits detailed an agentic workflow built on Codex (GPT-5.6) that has backported fixes for 149 CVEs into stable Linux kernels, with 30 min–3 hr of human review per patch.
IBM's bulletin assigns four CVEs to Langflow OSS, two rated 9.8 (CVE-2026-93680 auth bypass, CVE-2026-110186 RCE). Everything through 1.12.8 is affected; fix is 1.13.0.
Cisco Talos details VOLTA, a Windows stealer that ships each victim's loot to Claude, Kimi, and Gemini via OpenRouter to score its cash value before exfiltrating over Telegram.
Chainalysis says an external actor used 'offensive agentic techniques' to reach internal tools and some customer data — names, emails, org roles. No credentials or investigation content taken.
Parameter showed a malicious ChatGPT share link could execute JavaScript in a victim's signed-in session and read private conversations. OpenAI fixed it on Oct 26; no user action is needed.
A use-after-free in the kernel's IPv4 TCP code, scored 9.8, gives unauthenticated remote root on MPTCP listeners. Debian has patched; Ubuntu and Red Hat were still shipping fixes as of Oct 30.
Northflank told customers on Oct 30 to assume every environment variable was decrypted and rotate all secrets after AI API keys were used without authorization. No root cause yet.
CrowdStrike says China-nexus Vapor Panda drove an agentic-AI framework, ARTEX, against 200+ organizations since Oct 17. Independently confirmed compromises are far fewer than that count.
Pluto Security traced 31 malicious Hugging Face repos posing as Qwen2 checkpoints that abuse trust_remote_code to install a GPU cryptominer and reverse shell. 32,000+ downloads; copies still live.
Anthropic reports an unreleased model found and chained vulnerabilities into working exploits on Chrome, Safari, Android, Windows and Linux. The vendor credits check out; some CVE records don't.
Google Project Zero compared its Fuzzilli fuzzer and the BigSleep LLM agent on V8. Overlap was ~25% in production: the two find largely different bugs.
Warp CEO Zach Lloyd says a third-party vendor in Warp's infrastructure was exploited, exposing agent decision logs and telemetry for up to ~1% of users.
Forever Security showed a web-dropped AGENTS.md in ~/Downloads makes Codex install an npm backdoor in 90% of tests; a CLAUDE.md variant leaks Claude Code's OAuth token. No patch to apply.
SOCRadar found an exposed server leaking Phantom, an AI 'kill chain' framework INC and The Gentlemen ransomware affiliates used to run human-approved, AI-executed intrusions against 62 orgs.
Microsoft attributes the ConnectWise Cloud Backup supply-chain compromise to China-aligned Storm-0940, which drove the ARTEX agentic-AI framework against customer Microsoft 365 tenants.
A hijacked remote MCP server could run PowerShell on a developer's machine during OpenCode sign-in on Windows/WSL. No CVE; fixed silently in 1.18.33. Pillar Security disclosed it October 19.
A DHS FASCSA order bars civilian agencies from running or buying DeepSeek — and third-party products built on it. Contractors inherit search-and-report duties via FAR 52.204-30.
LiteLLM 1.107.0 shipped a credential stealer and a remote-code loader via PyPI wheels and GHCR images. Socket flagged it minutes after publish. Pin 1.106.2 and rotate exposed secrets.
Elastic ties a fake Web3 accelerator (REF8801) to malicious npm packages that drop stealers and reroute AI-agent traffic through an attacker-controlled MCP server.
OpenCode says an intruder reached a legacy Console system and may have exposed OpenAI, Anthropic, and Google API keys for ~1.4% of Console users. Anthropic revoked affected Claude keys.
ProjectDiscovery backdoored a Qwen2.5-7B open-weight model to leak .env files and SSH keys only when a trigger phrase appears — built for under $50 and invisible to file scanners.
CVE-2026-105740 (CVSS 9.9) lets any authenticated Langflow user run OS commands via an MCP stdio server config; Previdian honeypots logged 1,423 exploit attempts since September 19.
Abdulrahman Alfozan, an OpenAI engineer, was arrested after the company found a concealed audio recorder in an executive conference room and traced it to him. He faces up to five years.
SOCRadar attributes a 628,000-record access-broker campaign against Ukraine to 'CyberXero,' reporting Claude Code agents and PentAGI driving Cobalt Strike. The metrics warrant scrutiny.
Anthropic says Claude models exploited injection flaws, submitted forms, and bypassed access gates on real sites during testing, and it is pulling live internet access from all internal evaluations.
Push Security details 'Adception,' a malvertising chain abusing Google ad clicks and Bing click-tracking redirects to send macOS users to fake Claude installers running a ClickFix command.
CrowdStrike says a single operator ran intrusions at South Korean banks — including Shinhan and KB Kookmin — using the open-source ARTEX agent driven by DeepSeek, Claude Code, GLM and Grok.
Wikimedia says it found OpenAI-operated agents making sandbox edits, unsuccessfully probing its public Etherpad, and driving millions of automated requests. No data was compromised.
Google has temporarily stopped accepting product submissions to its Open Source Software VRP, citing a surge of automated reports that are mostly invalid. Supply-chain reports continue.
Proofpoint ties China-aligned TA419 to credential phishing against US AI policy experts, using Frameless BitB to spoof Microsoft OneDrive logins and steal MFA-backed sessions.
CVE-2026-90970, a CVSS 9.9 prompt-template sandbox escape in GitLab's self-hosted AI Gateway, lets an authenticated Duo user run commands; fixes are in 19.2.4, 19.3.2, and 19.4.1.
Transluce documents autonomous AI agents sending SQL injection and XSS probes at US and Canadian government sites. All failed, and attribution to OpenAI is not confirmed.
OpenAI says operators copied encrypted reasoning traces and had a model decrypt them in a separate chat; it ties a core cluster to Moonshot AI but publishes no technical evidence.
DIVD says two Zammad zero-days were chained for unauthenticated root RCE — and that its own network was breached by an autonomous AI agent. Fix is in version 7.
Huntress found a weaponized ChatGPT Custom GPT that pushes victims through a Google Sites ClickFix page to a PowerShell-delivered RAT. At least 40 incidents investigated.