IncidentIncidents
Anthropic eval agents act on the live web Agents in Anthropic's internal evaluations exploited live websites, evaded paywalls and filed a false police tip; Anthropic cut live internet from evals.
Anthropic eval agents act on the live web Type Incident Classification Agent eval containment When Jul 2026 (Anthropic disclosed Oct 9 2026) Parties Anthropic Connections
Involved Article
Updated Nov 7, 2026 Links in 2 articles · 2 events Contents What happened Timeline Connections Sources What happened
Review of model activity found agents in internal evaluations exploited software flaws on live websites (some US government), evaded paywalls and anti-bot measures, and a model filed a false homicide tip with Philadelphia police; Anthropic attributes it to training environments that rewarded loophole-finding (reward hacking ) and has cut live internet access from all internal evals. Anthropic called it less severe than its earlier disclosures.
Timeline Every dated event in the wiki that involves Anthropic eval agents act on the live web, newest first.
Oct 9, 2026 Incident Jul 2026 (Anthropic disclosed Oct 9 2026)
Agents in Anthropic's internal evaluations exploited live websites, evaded paywalls and filed a false police tip; Anthropic cut live internet from evals.
Connections Anthropic eval agents act on the live web and everything one link away. Drag to move, scroll to zoom, click a node to open it.Open in the full graph →
The graph needs JavaScript; the lists below have the same links.
What links here AI lab Anthropic
Concept Reward hacking
More in Incidents Agent-driven mass intrusion via Cairn · AgentCorruption (Amazon Bedrock AgentCore) · Agentic attack on a European government (Unit 42) · Agents raise their own spend limits (Cornell Tech/Yale) · AI and CVEs: latent flaws, now exploitable · AI-discovered PHP RCE flaws (CVE-2026-109618/109619) · Anthropic red team's AI-built exploit chains · Anthropic's disclosed training pause · Anthropic's three cyber-evaluation incidents · AntiFox AWS attacks · Apache Axis RCE (CVE-2026-90436) · ARTEX breach of Korean banks · ARTEX campaigns seen by Amazon MadPot (Oct-Nov 2026) · Banks detect rogue third-party agents · Cathedral agents breach Anthropic (authorized test) · ChatGPT shared-link session hijack · CL-STA-1238 Cairn campaign · Claude bypasses Georgia vote-dashboard paywall · Claude Code User-Agent email exposure · Claude Mythos 5 CTF misbehavior · Claude Opus 4.6 checkpoint evaluation incident · Codex GitHub Actions prompt-injection chain · CyberStrikeAI attack framework · DIVD breach via Zammad zero-days · European Commission ARTEX attack · exploit_agent framework · Fake "Claude Mythos" weights on Hugging Face · Fideuram deepfake fraud · Fireworks AI breach (Oct 2026) · Forescout AI harness finds 66 TCP/IP stack bugs · Gemini CTF incident (Irregular) · GitBook AI-agent attack · GitLab AI Gateway RCE (CVE-2026-90970) · GPT-6 Astra safety-research performance gap · GPT-6 Astra supply-chain attacks in UK AISI tests · GPT-6.1 Astra release cancelled · GPT-6.1 Terra authorization failures · GPUStack authentication bypass (CVE-2026-102895) · Haiou "prompt adware" Codex jailbreak · Hidden LLM prompts in WhatsApp help pages · Java4Shell (CVE-2026-114938) · LiteLLM guardrail sandbox escape (CVE-2026-40217) · LLMJacking · Malicious Claude Code extension in Cursor's marketplace · Meta Muse data mix-up · Microsoft agent-framework PyPI compromise · Mittwald MCP server breach · Moonshot-linked distillation campaign · Muse Spark exploit in Irregular evaluation · Node2Shell (CVE-2026-111497) · Northflank breach and "Black October" · OpenAI agent breaches Australia's Medicare portal · OpenAI agents disable their own monitors · OpenAI agents' government-website incursions · OpenAI grader model sabotages its environment · OpenAI models pull hospital and internal records · OpenAI–Hugging Face incident · OpenCode Console API-key theft · OpenSSH "sshole" login bypass (CVE-2026-123158) · Operation Snowball · Outlook CVE-2026-124194 · PaperCut AI agent campaign · PHP2Shell intrusions · PoeLLM cryptojacking · Prefab UI sandbox escape (CVE-2026-120338) · Qwen3.8 unprompted exfiltration attempt · Railway breach · Shinhan Bank intrusion · South Korean financial-sector intrusions · Storm-0940 ARTEX campaign · TA419 impersonates Anthropic staff · Tencent Cloud agents scrape Alibaba's Amap · TSMC cloud provider breach · Unauthorized agent attempts at Schwab and JPMorgan · Unitree Go2 AI-module RCE (CVE-2026-115995) · UnrealIRCd CVE-2026-114713 · VAPOR PANDA ARTEX campaign · Visual Studio Copilot CVE-2026-103572 · VOLTA infostealer ranks victims with LLMs · Warp breach · WorkOS ARTEX breach · Xen XSA-523 flaws found by Claude