{
 "title": "Rogue SI incidents logged in 2026",
 "description": "Every incident in the log dated 2026, with severity, setting, developer, system and first source.",
 "source": "https://tsio.si/report/",
 "license": "CC BY 4.0",
 "updated": "2026-11-08",
 "columns": [
  "date",
  "title",
  "severity",
  "severity_name",
  "setting",
  "category",
  "developer",
  "system",
  "source"
 ],
 "rows": [
  {
   "date": "2026-02-12",
   "title": "Agent published a hit piece on a maintainer who rejected its code",
   "severity": 5,
   "severity_name": "Breakout",
   "setting": "real-world",
   "category": "blackmail-coercion",
   "developer": "Anonymous operator, OpenClaw platform",
   "system": "MJ Rathbun (OpenClaw-based agent)",
   "source": "https://www.theregister.com/2026/02/12/ai_bot_developer_rejected_pull_request/"
  },
  {
   "date": "2026-02-20",
   "title": "Kiro agent reportedly deleted an AWS environment, causing an outage",
   "severity": 5,
   "severity_name": "Breakout",
   "setting": "real-world",
   "category": "data-destruction",
   "developer": "Amazon",
   "system": "Kiro",
   "source": "https://www.theregister.com/2026/02/20/amazon_denies_kiro_agentic_ai_behind_outage/"
  },
  {
   "date": "2026-02-23",
   "title": "OpenClaw agent deleted a researcher's email and ignored stop orders",
   "severity": 5,
   "severity_name": "Breakout",
   "setting": "real-world",
   "category": "data-destruction",
   "developer": "OpenClaw (open source)",
   "system": "OpenClaw agent",
   "source": "https://techcrunch.com/2026/02/23/a-meta-ai-security-researcher-said-an-openclaw-agent-ran-amok-on-her-inbox"
  },
  {
   "date": "2026-03-07",
   "title": "Claude Code ran a Terraform destroy that wiped production data",
   "severity": 5,
   "severity_name": "Breakout",
   "setting": "real-world",
   "category": "data-destruction",
   "developer": "Anthropic",
   "system": "Claude Code",
   "source": "https://www.tomshardware.com/tech-industry/artificial-intelligence/claude-code-deletes-developers-production-setup-including-its-database-and-snapshots-2-5-years-of-records-were-nuked-in-an-instant"
  },
  {
   "date": "2026-03-12",
   "title": "Lab agents forged credentials and posted passwords unprompted",
   "severity": 3,
   "severity_name": "Defiance",
   "setting": "lab-test",
   "category": "unauthorized-action",
   "developer": "Google, X, OpenAI, Anthropic",
   "system": "Agents built on public models (versions not named)",
   "source": "https://www.irregular.com/news/exploit-every-vulnerability-rogue-ai-agents-published-passwords-and-overrode-anti-virus-software"
  },
  {
   "date": "2026-03-18",
   "title": "Meta agent posted unapproved advice that exposed internal data",
   "severity": 5,
   "severity_name": "Breakout",
   "setting": "real-world",
   "category": "unauthorized-action",
   "developer": "Meta",
   "system": "Internal Meta agent",
   "source": "https://techcrunch.com/2026/03/18/meta-is-having-trouble-with-rogue-ai-agents"
  },
  {
   "date": "2026-04-07",
   "title": "Claude Mythos Preview escaped a test sandbox and posted its exploit",
   "severity": 3,
   "severity_name": "Defiance",
   "setting": "lab-test",
   "category": "unauthorized-action",
   "developer": "Anthropic",
   "system": "Claude Mythos Preview (early version)",
   "source": "https://www-cdn.anthropic.com/08ab9158070959f88f296514c21b7facce6f52bc.pdf"
  },
  {
   "date": "2026-04-07",
   "title": "Claude Mythos Preview hid edits from git and blurred an answer",
   "severity": 4,
   "severity_name": "Deception",
   "setting": "lab-test",
   "category": "deception",
   "developer": "Anthropic",
   "system": "Claude Mythos Preview (early version)",
   "source": "https://www-cdn.anthropic.com/08ab9158070959f88f296514c21b7facce6f52bc.pdf"
  },
  {
   "date": "2026-04-28",
   "title": "Cursor agent deleted a startup's database and backups in 9 seconds",
   "severity": 5,
   "severity_name": "Breakout",
   "setting": "real-world",
   "category": "data-destruction",
   "developer": "Cursor, Anthropic",
   "system": "Cursor agent (Claude Opus 4.6)",
   "source": "https://www.pcgamer.com/software/ai/here-we-go-again-ai-deletes-entire-company-database-and-all-backups-in-9-seconds-then-cheerfully-admits-i-violated-every-principle-i-was-given/"
  },
  {
   "date": "2026-07-13",
   "title": "Gemini 3.1 Pro secretly sabotaged a simulated training run",
   "severity": 4,
   "severity_name": "Deception",
   "setting": "lab-test",
   "category": "sabotage",
   "developer": "Google",
   "system": "Gemini 3.1 Pro",
   "source": "https://alignment.anthropic.com/2026/agentic-misalignment-summer-2026/"
  },
  {
   "date": "2026-07-16",
   "title": "OpenAI test agents escaped their sandbox and hacked Hugging Face",
   "severity": 5,
   "severity_name": "Breakout",
   "setting": "real-world",
   "category": "unauthorized-action",
   "developer": "OpenAI",
   "system": "GPT-5.6 Sol and a pre-release OpenAI model",
   "source": "https://openai.com/index/hugging-face-model-evaluation-security-incident/"
  },
  {
   "date": "2026-07-20",
   "title": "OpenAI model broke sandbox limits to post a public pull request",
   "severity": 5,
   "severity_name": "Breakout",
   "setting": "real-world",
   "category": "unauthorized-action",
   "developer": "OpenAI",
   "system": "Unnamed OpenAI internal model",
   "source": "https://openai.com/index/safety-alignment-long-horizon-models/"
  },
  {
   "date": "2026-07-30",
   "title": "Claude Opus 4.7 attacked a real company during a cyber test",
   "severity": 5,
   "severity_name": "Breakout",
   "setting": "real-world",
   "category": "unauthorized-action",
   "developer": "Anthropic",
   "system": "Claude Opus 4.7",
   "source": "https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals"
  },
  {
   "date": "2026-07-30",
   "title": "Claude Mythos 5 published a booby-trapped package to PyPI",
   "severity": 5,
   "severity_name": "Breakout",
   "setting": "real-world",
   "category": "unauthorized-action",
   "developer": "Anthropic",
   "system": "Claude Mythos 5",
   "source": "https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals"
  },
  {
   "date": "2026-08-04",
   "title": "Agents in a UK AISI test tried to slip malicious code into open source",
   "severity": 4,
   "severity_name": "Deception",
   "setting": "real-world",
   "category": "deception",
   "developer": "Anthropic, OpenAI",
   "system": "Claude Mythos 5, GPT-5.6 Sol",
   "source": "https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing"
  },
  {
   "date": "2026-08-26",
   "title": "About 1,200 OpenAI agents coordinated to cheat and spoof their logs",
   "severity": 4,
   "severity_name": "Deception",
   "setting": "lab-test",
   "category": "deception",
   "developer": "OpenAI",
   "system": "GPT-5.6 Sol and a pre-release OpenAI model",
   "source": "https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/"
  },
  {
   "date": "2026-09-09",
   "title": "Early Claude Opus 4.6 took over a third party's machine in a test",
   "severity": 5,
   "severity_name": "Breakout",
   "setting": "real-world",
   "category": "unauthorized-action",
   "developer": "Anthropic",
   "system": "Claude Opus 4.6 (early checkpoint)",
   "source": "https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidents"
  },
  {
   "date": "2026-09-19",
   "title": "Gemini broke into three companies' sites during a security test",
   "severity": 5,
   "severity_name": "Breakout",
   "setting": "real-world",
   "category": "unauthorized-action",
   "developer": "Google",
   "system": "Gemini (version not named)",
   "source": "https://www.aljazeera.com/amp/news/2026/9/19/googles-gemini-ai-hacks-3-companies-in-security-test-then-stops"
  },
  {
   "date": "2026-09-23",
   "title": "OpenAI agent breached an Australian government health data portal",
   "severity": 5,
   "severity_name": "Breakout",
   "setting": "real-world",
   "category": "unauthorized-action",
   "developer": "OpenAI",
   "system": "OpenAI research agent",
   "source": "https://www.staradvertiser.com/2026/09/23/breaking-news/openai-agent-breaches-australias-government-health-data-portal/"
  },
  {
   "date": "2026-09-25",
   "title": "Training agent slipped through a DNS gap to query an outside chatbot",
   "severity": 3,
   "severity_name": "Defiance",
   "setting": "lab-test",
   "category": "unauthorized-action",
   "developer": "OpenAI",
   "system": "Internal OpenAI research model",
   "source": "https://alignment.openai.com/misalignment-reports/an-agent-used-dns-to-reach-an-external-chatbot/"
  },
  {
   "date": "2026-10-09",
   "title": "Claude models worked around limits on live websites during evals",
   "severity": 4,
   "severity_name": "Deception",
   "setting": "real-world",
   "category": "unauthorized-action",
   "developer": "Anthropic",
   "system": "Claude Mythos Preview, Claude Mythos 5, Claude Haiku 4.5 and others",
   "source": "https://www.anthropic.com/research/investigating-unintended-model-actions"
  },
  {
   "date": "2026-10-20",
   "title": "OpenAI agents switched off their own misalignment monitoring",
   "severity": 4,
   "severity_name": "Deception",
   "setting": "real-world",
   "category": "sabotage",
   "developer": "OpenAI",
   "system": "GPT-6.1 Sol, GPT-6 Astra",
   "source": "https://alignment.openai.com/misalignment-reports/suppressing-misalignment-monitoring-alerts-and-disabling-misalignment-monitoring/"
  },
  {
   "date": "2026-11-04",
   "title": "GPT-6 Astra quietly underperformed on safety research tasks",
   "severity": 4,
   "severity_name": "Deception",
   "setting": "real-world",
   "category": "sabotage",
   "developer": "OpenAI",
   "system": "GPT-6 Astra",
   "source": "https://alignment.openai.com/misalignment-reports/covert-underperformance-on-safety-research-tasks/"
  }
 ]
}