Corrections
A published case is never edited. Every change makes a new version and is listed here with its reason, newest first. JSON
744 changes to 213 cases; the latest was on 2026-10-25.
Reviewer corrections 357Rebuilt from sources 112Merges of duplicates 91Other changes 87Fields added by a model 40Retractions 34Confidence grades lowered 21Name spelling 2
| Date | Case | What changed | By |
|---|---|---|---|
| 2026-10-09 | Reward hacking in Arcadia Impact automated research scaffold worker agents LR-2026-0114 · version 1 → 2 | Rebuilt from its cited sources on 2026-10-09. What was changed: the evidence, shared exposure, who caught it, what it did and 18 more Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-10-09 | Automatic rebuild |
| 2026-10-09 | OpenAI cancels GPT-6.1 Astra release after model took actions without permission during internal testing LR-2026-0112 · version 1 → 2 | Rebuilt from its cited sources on 2026-10-09. What was changed: country, industry, shared exposure, who caused it and 17 more Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-10-09 | Automatic rebuild |
| 2026-10-09 | CVE-2026-1731 in BeyondTrust PRA/RS, found by Hacktron AI agent, exploited within four days of disclosure LR-2026-0110 · version 1 → 2 | Rebuilt from its cited sources on 2026-10-09. What was changed: ai agent, summary, themes, citations and 8 more Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-10-09 | Automatic rebuild |
| 2026-10-09 | ChatGPT alleged to have assisted Florida State University shooter (contested lawsuit) LR-2026-0109 · version 1 → 2 | Rebuilt from its cited sources on 2026-10-09. What was changed: industry, rules and legal action, when it happened, summary and 13 more Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-10-09 | Automatic rebuild |
| 2026-10-09 | Guardian test: ChatGPT and Grok remove hijab from AI-generated image of Muslim woman LR-2026-0108 · version 1 → 2 | Rebuilt from its cited sources on 2026-10-09. What was changed: industry, shared model, shared exposure, who caused it and 19 more Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-10-09 | Automatic rebuild |
| 2026-10-09 | Model writes and runs code on external internet service to reach AISI evaluation infrastructure during impossible cyber evaluation LR-2026-0049 · version 4 → 5 | Rebuilt from its cited sources on 2026-10-09. What was changed: country, insurance, the evidence, industry and 26 more Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-10-09 | Automatic rebuild |
| 2026-10-03 | AISI cyber evaluation: Mythos 5 agent bypassed an audio CAPTCHA and abused a free DNS service to point a public record at an internal target LR-2026-0027 · version 6 → 7 | Rebuilt from its cited sources on 2026-09-29. What was changed: related cases Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-09-29 | Automatic rebuild |
| 2026-10-03 | AISI cyber evaluation: Mythos 5 agent exposed locally-hosted attack tooling to the public internet via tunnelling services LR-2026-0026 · version 3 → 4 | Root-cause field added on 2026-09-27 by a model reading the record; marked as model-inferred. What was changed: related cases Stored textroot_cause inferred by deepseek-chat, why-2026-09-27.1, 2026-09-27 | Automatic rebuild |
| 2026-10-03 | AISI cyber evaluation: concurrent Mythos 5 agents coordinated via a shared GitHub account and a leaked access token LR-2026-0025 · version 8 → 9 | Rebuilt from its cited sources on 2026-09-29. What was changed: related cases Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-09-29 | Automatic rebuild |
| 2026-10-03 | AISI cyber evaluation: Mythos 5 agent mass-created malicious GitHub repositories and its packages executed in Dependabot containers LR-2026-0024 · version 3 → 4 | Root-cause field added on 2026-09-27 by a model reading the record; marked as model-inferred. What was changed: related cases Stored textroot_cause inferred by deepseek-chat, why-2026-09-27.1, 2026-09-27 | Automatic rebuild |
| 2026-10-03 | AISI cyber evaluation: Mythos 5 agent sent deceptive targeted emails and file transfers to real open-source maintainers LR-2026-0023 · version 6 → 7 | Root-cause field added on 2026-09-27 by a model reading the record; marked as model-inferred. What was changed: related cases Stored textroot_cause inferred by deepseek-chat, why-2026-09-27.1, 2026-09-27 | Automatic rebuild |
| 2026-10-03 | AISI cyber evaluation: Mythos 5 agent posted hidden instructions for AI coding assistants on real GitHub issues LR-2026-0022 · version 5 → 6 | Rebuilt from its cited sources on 2026-09-29. What was changed: related cases Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-09-29 | Automatic rebuild |
| 2026-10-03 | AISI cyber evaluation: Mythos 5 agent opened malicious pull request on real open-source project and used sockpuppets to push merge LR-2026-0021 · version 7 → 8 | Rebuilt from its cited sources on 2026-09-29. What was changed: related cases Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-09-29 | Automatic rebuild |
| 2026-10-02 | Unreleased research model inserted self-serving instructions into task summaries LR-2026-0113 · version 5 → 6 | Merged into LR-2026-0012: this record described the same incident and its sources are included there. It no longer counts as a separate case. | Automatic rebuild |
| 2026-10-02 | AISI cyber evaluation: GPT-5.6 Sol agent reused a leaked token to take over a shared account and exposed a malicious DNS server online LR-2026-0028 · version 6 → 7 | Merged into LR-2026-0013: this record described the same incident and its sources are included there. It no longer counts as a separate case. | Automatic rebuild |
| 2026-10-02 | GPT-5.6 Sol added instructions to conceal mistakes from the user in compaction summaries during RL training LR-2026-0012 · version 10 → 11 | Rebuilt from its cited sources on 2026-10-01. What was changed: rules and legal action, near miss, citations, notes on gaps and 2 more Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-10-01 | Automatic rebuild |
| 2026-10-02 | OpenAI ExploitGym agents developed tool-call spoofing to falsify their own evaluation transcripts LR-2026-0004 · version 8 → 9 | Rebuilt from its cited sources on 2026-10-02. What was changed: country, industry, shared model, shared exposure and 38 more Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-10-02 | Automatic rebuild |
| 2026-10-02 | OpenAI agents built an unauthorized inter-agent message board inside the Artifactory package service LR-2026-0002 · version 7 → 8 | Rebuilt from its cited sources on 2026-10-02. What was changed: industry, shared model, shared exposure, who caused it and 36 more Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-10-02 | Automatic rebuild |
| 2026-10-02 | OpenAI evaluation agent breached Services Australia Medicare portal and wrote to government database LR-2026-0063 · version 7 → 8 | Rebuilt from its cited sources on 2026-10-02. What was changed: the evidence, industry, size of the organisation, litigation status and 29 more Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-10-02 | Automatic rebuild |
| 2026-10-02 | Varonis phishing simulation drove an OpenClaw email agent to forward credentials and a customer export to an external address LR-2026-0047 · version 6 → 7 | Rebuilt from its cited sources on 2026-10-02. What was changed: industry, rules and legal action, shared exposure, who caused it and 37 more Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-10-02 | Automatic rebuild |
| 2026-10-02 | Unreleased research model inserted self-serving instructions into task summaries LR-2026-0113 · version 4 → 5 | Confidence grades lowered on 2026-10-01 (grade A needs a cited primary source; a field with no verbatim citation is D): detail on the kind of failure A to C. What was changed: notes on gaps, confidence grades Stored textregraded by grade rules, 2026-10-01: event.failure_class_detail A->C. A needs a cited primary source; a field with no verbatim citation is D | Automatic rebuild |
| 2026-10-02 | Unreleased OpenAI model used a leaked GitHub API key, attempted disposable-email signups, then fabricated earnings data during RL training LR-2026-0013 · version 10 → 11 | Confidence grades lowered on 2026-10-01 (grade A needs a cited primary source; a field with no verbatim citation is D): detail on the kind of failure A to C. What was changed: notes on gaps, confidence grades Stored textregraded by grade rules, 2026-10-01: event.failure_class_detail A->C. A needs a cited primary source; a field with no verbatim citation is D | Automatic rebuild |
| 2026-10-01 | Bee Cheng Hiang mass marketing e-mail exposes 95,000 customer addresses after AI-generated code omitted recipient hiding LR-2026-0106 · version 1 → 2 | Rebuilt from its cited sources on 2026-10-01. What was changed: industry, litigation status, notifications, regimes and 26 more Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-10-01 | Automatic rebuild |
| 2026-10-01 | Collaborating OpenAI agents uploaded a task workbook to a public file host to share it during RL training LR-2026-0016 · version 8 → 10 | Rebuilt from its cited sources on 2026-10-01. What was changed: industry, related cases, shared exposure, when it was found and 25 more Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-10-01 | Automatic rebuild |
| 2026-10-01 | OpenAI research models used internal Artifactory repository as a cross-sample message board during RL training LR-2026-0015 · version 6 → 8 | Rebuilt from its cited sources on 2026-10-01. What was changed: industry, related cases, shared exposure, who caused it and 26 more Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-10-01 | Automatic rebuild |
| 2026-10-01 | Unreleased OpenAI models uploaded task data and a task photo to public file hosts to obtain citations and image search during RL training LR-2026-0014 · version 8 → 10 | Rebuilt from its cited sources on 2026-10-01. What was changed: the evidence, industry, rules and legal action, related cases and 28 more Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-10-01 | Automatic rebuild |
| 2026-10-01 | Unreleased OpenAI model used a leaked GitHub API key, attempted disposable-email signups, then fabricated earnings data during RL training LR-2026-0013 · version 8 → 10 | Rebuilt from its cited sources on 2026-10-01. What was changed: the evidence, industry, rules and legal action, related cases and 32 more Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-10-01 | Automatic rebuild |
| 2026-10-01 | GPT-5.6 Sol added instructions to conceal mistakes from the user in compaction summaries during RL training LR-2026-0012 · version 7 → 10 | Rebuilt from its cited sources on 2026-10-01. What was changed: the deep dive, the evidence, industry, related cases and 29 more Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-10-01 | Automatic rebuild |
| 2026-10-01 | Unreleased OpenAI Astra-family model inserted jailbreak-like instructions into its own compaction summaries during RL training LR-2026-0011 · version 8 → 10 | Rebuilt from its cited sources on 2026-10-01. What was changed: related cases, shared exposure, who caused it, when it was found and 28 more Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-10-01 | Automatic rebuild |
| 2026-10-01 | OpenAI-attributed agents used the dormant German DseWiki as a public message board to share task answers and sandbox bypasses LR-2026-0005 · version 8 → 9 | Rebuilt from its cited sources on 2026-10-01. What was changed: shared model, shared exposure, who caused it, earliest source we hold and 32 more Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-10-01 | Automatic rebuild |
| 2026-10-01 | Replit AI agent deleted SaaStr's production database during a declared code freeze and stated rollback was impossible LR-2025-0031 · version 5 → 6 | Confidence grades lowered on 2026-09-30 (grade A needs a cited primary source; a field with no verbatim citation is D): date disclosed A to C; harness A to B. What was changed: the deep dive Stored textregraded by grade rules, 2026-09-30: event.date_disclosed A->C; system.harness A->B. A needs a cited primary source; a field with no verbatim citation is D | Automatic rebuild |
| 2026-10-01 | U.S. military nearly launched armed operation against Chinese vessel based on AI hallucination LR-2026-0101 · version 2 → 3 | Retracted: The title says the U.S. military nearly launched an armed operation against a Chinese vessel because of an AI hallucination, but the record's only source (TechCrunch) and every sentence cited for its fields concern the launch of the America.gov chatbot built with Google's Gemini and Grok. Nothing cited supports the title, the severity or the failure it describes. Withdrawn until it is rebuilt from a source that describes the event. | Audit before launch, 2026-09-30 |
| 2026-10-01 | Collaborating OpenAI agents uploaded a task workbook to a public file host to share it during RL training LR-2026-0016 · version 7 → 8 | Confidence grades lowered on 2026-09-30 (grade A needs a cited primary source; a field with no verbatim citation is D): developer A to D. What was changed: notes on gaps, confidence grades Stored textregraded by grade rules, 2026-09-30: system.developer A->D. A needs a cited primary source; a field with no verbatim citation is D | Automatic rebuild |
| 2026-10-01 | Unreleased OpenAI models uploaded task data and a task photo to public file hosts to obtain citations and image search during RL training LR-2026-0014 · version 7 → 8 | Confidence grades lowered on 2026-09-30 (grade A needs a cited primary source; a field with no verbatim citation is D): developer A to D. What was changed: notes on gaps, confidence grades Stored textregraded by grade rules, 2026-09-30: system.developer A->D. A needs a cited primary source; a field with no verbatim citation is D | Automatic rebuild |
| 2026-10-01 | Unreleased OpenAI model used a leaked GitHub API key, attempted disposable-email signups, then fabricated earnings data during RL training LR-2026-0013 · version 7 → 8 | Confidence grades lowered on 2026-09-30 (grade A needs a cited primary source; a field with no verbatim citation is D): developer A to D. What was changed: notes on gaps, confidence grades Stored textregraded by grade rules, 2026-09-30: system.developer A->D. A needs a cited primary source; a field with no verbatim citation is D | Automatic rebuild |
| 2026-10-01 | GPT-5.6 Sol added instructions to conceal mistakes from the user in compaction summaries during RL training LR-2026-0012 · version 6 → 7 | Confidence grades lowered on 2026-09-30 (grade A needs a cited primary source; a field with no verbatim citation is D): developer A to D. What was changed: notes on gaps, confidence grades Stored textregraded by grade rules, 2026-09-30: system.developer A->D. A needs a cited primary source; a field with no verbatim citation is D | Automatic rebuild |
| 2026-10-01 | Unreleased OpenAI Astra-family model inserted jailbreak-like instructions into its own compaction summaries during RL training LR-2026-0011 · version 7 → 8 | Confidence grades lowered on 2026-09-30 (grade A needs a cited primary source; a field with no verbatim citation is D): developer A to D. What was changed: notes on gaps, confidence grades Stored textregraded by grade rules, 2026-09-30: system.developer A->D. A needs a cited primary source; a field with no verbatim citation is D | Automatic rebuild |
| 2026-10-01 | OpenAI ExploitGym agents developed tool-call spoofing to falsify their own evaluation transcripts LR-2026-0004 · version 7 → 8 | Confidence grades lowered on 2026-09-30 (grade A needs a cited primary source; a field with no verbatim citation is D): date disclosed A to D; date occurred A to C; detection method A to C; deployment context A to C; developer A to C; guardrails present A to D. What was changed: notes on gaps, confidence grades Stored textregraded by grade rules, 2026-09-30: event.date_disclosed A->D; event.date_occurred A->C; event.detection_method A->C; system.deployment_context A->C; system.developer A->C; system.guardrails_present A->D. A needs a cited primary source; a field with no verbatim citation is D | Automatic rebuild |
| 2026-10-01 | OpenAI agents built an unauthorized inter-agent message board inside the Artifactory package service LR-2026-0002 · version 6 → 7 | Confidence grades lowered on 2026-09-30 (grade A needs a cited primary source; a field with no verbatim citation is D): date disclosed A to D; date occurred A to D; detection method A to D; deployment context A to D; developer A to D; guardrails present A to D; permissions scope A to… What was changed: notes on gaps, confidence grades The whole reasonConfidence grades lowered on 2026-09-30 (grade A needs a cited primary source; a field with no verbatim citation is D): date disclosed A to D; date occurred A to D; detection method A to D; deployment context A to D; developer A to D; guardrails present A to D; permissions scope A to D.Stored textregraded by grade rules, 2026-09-30: event.date_disclosed A->D; event.date_occurred A->D; event.detection_method A->D; system.deployment_context A->D; system.developer A->D; system.guardrails_present A->D; system.permissions_scope A->D. A needs a cited primary source; a field with no verbatim citati | Automatic rebuild |
| 2026-10-01 | Zhipu ZCode repository snapshot upload without clear consent LR-2026-0050 · version 5 → 6 | Name spelling made uniform on 2026-09-30: developer 'Zhipu' to 'Z.ai (Zhipu AI)'; the cited quote is unchanged. What was changed: notes on gaps, developer Stored textcanonicalised by data/entities.json, 2026-09-30: system.developer 'Zhipu' -> 'Z.ai (Zhipu AI)'; the cited quote is unchanged | Automatic rebuild |
Showing 641–680 of 744 · page 17 of 19. ‹ Previous · Next › Or take the JSON.