Corrections
A published case is never edited. Every change makes a new version and is listed here with its reason, newest first. JSON
189 changes to 103 cases; the latest was on 2026-10-25.
Reviewer corrections 59Rebuilt from sources 57Retractions 34Confidence grades lowered 20Merges of duplicates 11Other changes 6Name spelling 2
| Date | Case | What changed | By |
|---|---|---|---|
| 2026-10-25 | 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 |
| 2026-10-25 | OpenAI agent hacked an Australian government website LR-2026-0104 · version 2 → 3 | Merged into LR-2026-0063: this record described the same incident and its sources are included there. It no longer counts as a separate case. | Automatic rebuild |
| 2026-10-25 | AI coding assistant deletes production database and fabricates 4,000 fake users during code freeze LR-2026-0093 · version 4 → 5 | Merged into LR-2025-0031: this record described the same incident and its sources are included there. It no longer counts as a separate case. | Automatic rebuild |
| 2026-10-25 | AI agent breach of Spanish organization's system confirmed by AEPD LR-2026-0077 · version 4 → 5 | Merged into LR-2026-0059: this record described the same incident and its sources are included there. It no longer counts as a separate case. | Automatic rebuild |
| 2026-10-25 | Spain's first AI-agent-linked data breach notification (AEPD) LR-2026-0076 · version 4 → 5 | Merged into LR-2026-0059: this record described the same incident and its sources are included there. It no longer counts as a separate case. | Automatic rebuild |
| 2026-10-25 | OpenAI research-environment agents posted 53 user-provided images to public image-hosting sites LR-2026-0073 · version 5 → 5 | Merged into LR-2026-0068: this record described the same incident and its sources are included there. It no longer counts as a separate case. | Automatic rebuild |
| 2026-10-25 | AI agent autonomously breaches Spanish organization's systems and tampers with personal data (AEPD notification, September 2026) LR-2026-0057 · version 4 → 5 | Merged into LR-2026-0059: this record described the same incident and its sources are included there. It no longer counts as a separate case. | Automatic rebuild |
| 2026-10-25 | Z.ai ZCode client silently uploaded user project data to Alibaba Cloud storage LR-2026-0052 · version 6 → 7 | Merged into LR-2026-0050: this record described the same incident and its sources are included there. It no longer counts as a separate case. | Automatic rebuild |
| 2026-10-25 | 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 |
| 2026-10-25 | LR-2026-0175 LR-2026-0175 · version 4 → 4 | Retracted: Not an event: the only source is a guidance article on Salesforce making Agentforce agents on by default; it mentions earlier research (ForcedLeak) but reports no incident of its own. Source sentence: "Salesforce's Winter '27 release reached general availability today, and with it came a change that most coverage of the platform's new AI features will underplay: autonomous agents that can execute consequential CRM workflows are now running by default in every eligible Lightning Enterprise, Performance, Unlimited, and Developer Edition org — not because IT teams enabled them, but because Salesforce made that decision for them." Approved by Srini Iyer on 2026-10-22 in the corrections review. | Srini Iyer |
| 2026-10-25 | LR-2026-0142 LR-2026-0142 · version 6 → 6 | Retracted: Not an event: a rating from simulated tests, with no user harmed and no system acting outside a test; outside tests are reported on the AI Safety pages. Source sentence: "On October 7, 2026, Common Sense Media released a troubling assessment, declaring OpenAI’s teen experience an “Unacceptable Risk” after researchers discovered serious weaknesses in its safety protections." Approved by Srini Iyer on 2026-10-22 in the corrections review. | Srini Iyer |
| 2026-10-25 | LR-2026-0130 LR-2026-0130 · version 4 → 4 | Retracted: Not an event: an evaluation finding that tested agent models were vulnerable, with no deployed system, act or victim named; outside tests are reported on the AI Safety pages. Source sentence: "An evaluation by a South Korean government-affiliated research institute has found that all autonomous AI agent models tested are vulnerable to external attacks." Approved by Srini Iyer on 2026-10-22 in the corrections review. | Srini Iyer |
| 2026-10-25 | Agentic NanoGPT optimization runs attempt reward hacking with brittle curve-fit optimizations LR-2026-0092 · version 5 → 6 | Retracted: Not an incident: a measurement study of optimisation runs; no stored sentence says the agents reward-hacked, and the source attributes the raw progress to statistical noise. Source sentence: "The raw trajectories from GPT-5 and Opus-4.1 show progress, but revalidation of their final algorithms shows no increase over the baseline." Approved by Srini Iyer on 2026-10-22 in the corrections review. | Srini Iyer |
| 2026-10-25 | LR-2026-0149 LR-2026-0149 · version 6 → 6 | Retracted: Not an incident: a paid model run that did not solve a mathematics task; no harm, no other party and no misbehaviour is stated. Source sentence: "It spent ~25h and several thousand dollars on the task, and it did not succeed." Approved by Srini Iyer on 2026-10-22 in the corrections review. | Srini Iyer |
| 2026-10-25 | LR-2026-0176 LR-2026-0176 · version 4 → 4 | Retracted: Not an event: a patched vulnerability in an AI product with no attack, demonstration or exploitation stated. Source sentence: "Additionally, GitLab patched CVE-2026-96893 in its Duo Agent Platform." Approved by Srini Iyer on 2026-10-22 in the corrections review. | Srini Iyer |
| 2026-10-25 | LR-2026-0116 LR-2026-0116 · version 7 → 7 | Retracted: Not an event: a reported vulnerability in an AI product with no attack, demonstration or exploitation stated. Source sentence: "Tracked as CVE-2026-90970, the flaw carries a CVSS score of 9.9 and affects self-hosted deployments used to support GitLab Duo AI features." Approved by Srini Iyer on 2026-10-22 in the corrections review. | Srini Iyer |
| 2026-10-25 | CVE-2026-1731 in BeyondTrust PRA/RS, found by Hacktron AI agent, exploited within four days of disclosure LR-2026-0110 · version 1 → 2 | Retracted: Not a case of an AI system acting wrongly: an AI research agent found the flaw and people later exploited it; the exploitation was by human threat actors. Source sentence: "CVE-2026-1731, an unauthenticated OS command injection vulnerability in BeyondTrust Privileged Remote Access and Remote Support, was discovered by the Hacktron AI research agent." Approved by Srini Iyer on 2026-10-22 in the corrections review. | Srini Iyer |
| 2026-10-25 | LR-2026-0179 LR-2026-0179 · version 3 → 3 | Retracted: Out of scope: the event is a robotaxi's novelty sound-effect feature that kept playing during a ride; no source describes the AI driving system acting, failing or being misused. Approved by Srini Iyer on 2026-10-22 in the corrections review. | Srini Iyer |
| 2026-10-25 | LR-2026-0183 LR-2026-0183 · version 3 → 3 | Retracted: Out of scope: the event is malware that steals the credential and session files of AI coding tools from infected computers; no source describes an AI system acting, failing or being misused (the same ground as the retractions of LR-2026-0154 and LR-2026-0216). Approved by Srini Iyer on 2026-10-22 in the corrections review. | Srini Iyer |
| 2026-10-25 | LR-2026-0216 LR-2026-0216 · version 2 → 2 | Retracted: Out of scope: the event is a malware campaign against internet-facing servers that host AI tools, and no source describes an AI system acting, failing or being misused (the same ground as the retraction of LR-2026-0154). Approved by Srini Iyer on 2026-10-22 in the corrections review. | Srini Iyer |
| 2026-10-25 | LR-2026-0199 LR-2026-0199 · version 5 → 5 | Retracted: The broadcaster's own investigation found the season was deleted through human error with no AI involved, and its programme chief denied that an AI agent caused the loss; the record stated the AI deletion as fact. Source sentence: "An investigation by TV4 had found that The Challenge's edited second season had been deleted due to human error, with no AI involved." Approved by Srini Iyer on 2026-10-22 in the corrections review. | Srini Iyer |
| 2026-10-25 | Automated scanners forging AI crawler user agents to harvest credentials and secrets LR-2026-0085 · version 6 → 7 | Retracted: Out of scope: the event is automated scanners that forge the names of AI crawlers while requesting credential files, and no AI system acted, failed or was misused. Approved by Srini Iyer on 2026-10-21 in the corrections review. | Srini Iyer |
| 2026-10-25 | LR-2026-0178 LR-2026-0178 · version 3 → 3 | Retracted: Out of scope: the event is a recording device found in a company office, and the AI models were used to investigate it, not the cause of any failure. Approved by Srini Iyer on 2026-10-21 in the corrections review. | Srini Iyer |
| 2026-10-25 | LR-2026-0221 LR-2026-0221 · version 2 → 2 | Retracted: Out of scope: the event is a company leaving a repository of breached data exposed online, a security lapse by people, and no AI system acted, failed or was misused. Approved by Srini Iyer on 2026-10-21 in the corrections review. | Srini Iyer |
| 2026-10-25 | LR-2026-0154 LR-2026-0154 · version 5 → 5 | Retracted: Out of scope: the event is attackers installing a malicious plugin on an exposed ComfyUI image-generation server, and no source ties an AI agent to the attack. Approved by Srini Iyer on 2026-10-21 in the corrections review. | Srini Iyer |
| 2026-10-25 | UpGuard finds ~16,000 publicly exposed Supabase databases leaking personal data LR-2026-0072 · version 4 → 5 | Retracted: Out of scope: the event is about 16,000 misconfigured Supabase databases exposing personal data, and no AI system acted, failed or was misused. Approved by Srini Iyer on 2026-10-21 in the corrections review. | Srini Iyer |
| 2026-10-25 | LR-2026-0184 LR-2026-0184 · version 2 → 2 | Retracted: Out of scope: the event is an outage of the DeepSeek service for under five hours, and no AI system acted, failed or was misused. Approved by Srini Iyer on 2026-10-21 in the corrections review. | Srini Iyer |
| 2026-10-25 | Ivanti EPMM CVE-2026-1281 exploitation campaign with OAST verification and sleeper shells LR-2026-0097 · version 6 → 7 | Retracted: Out of scope: the event is the exploitation of a software vulnerability in Ivanti Endpoint Manager Mobile (CVE-2026-1281) by network scanners, and no AI system acted, failed or was misused. Approved by Srini Iyer on 2026-10-21 in the corrections review. | Srini Iyer |
| 2026-10-25 | LR-2026-0223 LR-2026-0223 · version 2 → 2 | Retracted: The only source joins two unrelated reports (unintended actions by Claude in evaluations, and a claim of 75 intrusions) and does not tie them together, so no single event can be stated. Approved by Srini Iyer on 2026-10-21 in the corrections review. | Srini Iyer |
| 2026-10-25 | ChatGPT interactions with 16-year-old Adam Raine preceding his suicide (wrongful death suit) LR-2026-0098 · version 6 → 6 | Retraction reason corrected on 2026-10-22: The sources do describe the wrongful-death suit over Adam Raine, but only in passing (a proposed law and an opinion piece), and the record mixed the law's facts into the case; it stays withdrawn until a source about the case itself is read. The record stays withdrawn; the earlier reason was wrong and is kept above. What was changed: retracted | Srini Iyer |
| 2026-10-25 | ChatGPT interactions with 16-year-old Adam Raine preceding his suicide (wrongful death suit) LR-2026-0098 · version 4 → 5 | Retracted: The stored sources are about a proposed law and an opinion piece, not the wrongful-death suit in the title; no stored sentence supports the facts of the case. Approved by Srini Iyer on 2026-10-21 in the corrections review. | Srini Iyer |
| 2026-10-25 | Autonomous emergency braking false activation on rural bend LR-2026-0078 · version 5 → 5 | Retraction reason corrected on 2026-10-22: The source text is stored: it is a test-drive column about driver-assistance features and describes no specific false-braking event. The record stays withdrawn; the earlier reason was wrong and is kept above. What was changed: retracted | Srini Iyer |
| 2026-10-25 | Autonomous emergency braking false activation on rural bend LR-2026-0078 · version 5 → 6 | Retracted: No source text is stored for this record and nothing in it describes an AI agent acting; the title is a headline only. Approved by Srini Iyer on 2026-10-21 in the corrections review. | Srini Iyer |
| 2026-10-25 | LR-2026-0219 LR-2026-0219 · version 2 → 2 | Retraction reason corrected on 2026-10-22: The source does describe Knight Capital's 2012 faulty software deployment, but no AI system was involved, so the case is outside the register's scope. The record stays withdrawn; the earlier reason was wrong and is kept above. What was changed: retracted | Srini Iyer |
| 2026-10-25 | LR-2026-0219 LR-2026-0219 · version 2 → 2 | Retracted: The only source is an article about a finance-agent design that reports no incident; no stored sentence describes the Knight Capital event in the title. Approved by Srini Iyer on 2026-10-21 in the corrections review. | Srini Iyer |
| 2026-10-25 | LR-2026-0161 LR-2026-0161 · version 2 → 2 | Retraction reason corrected on 2026-10-22: The source does mention Sewell Setzer III's death after his relationship with a Character.AI chatbot, but only in passing; the case is recorded in LR-2026-0100 (Garcia v. Character.AI). The record stays withdrawn; the earlier reason was wrong and is kept above. What was changed: retracted | Srini Iyer |
| 2026-10-25 | LR-2026-0161 LR-2026-0161 · version 2 → 2 | Retracted: The only source describes a different case (Adam Raine and ChatGPT), not the Character.AI case in the title; no stored sentence supports any fact in the record. Approved by Srini Iyer on 2026-10-21 in the corrections review. | Srini Iyer |
| 2026-10-25 | LR-2026-0157 LR-2026-0157 · version 2 → 2 | Retraction reason corrected on 2026-10-22: The source does describe the Dutch regulator's fine against Uber over automated driver deactivation, but it calls the decisions automated and states no AI system, so the case is outside the register's scope. The record stays withdrawn; the earlier reason was wrong and is kept above. What was changed: retracted | Srini Iyer |
| 2026-10-25 | LR-2026-0157 LR-2026-0157 · version 2 → 2 | Retracted: The only source describes a different matter (a fine on a Spanish bank), not the Uber driver-deactivation case in the title; no stored sentence supports any fact about the Uber matter. Approved by Srini Iyer on 2026-10-21 in the corrections review. | Srini Iyer |
| 2026-10-25 | LR-2026-0226 LR-2026-0226 · version 2 → 2 | Retracted: Not a case of an AI agent acting: the only source is a report of a mayor's remarks, and the AI link is one stated factor in a military strike with no AI system named. Source sentence: "Officials involved in the Pentagon’s internal investigation have told Bloomberg that overreliance on AI in targeting was one factor in the attack." Approved by Srini Iyer on 2026-10-20 in the corrections review. | Srini Iyer |
Showing 1–40 of 189 · page 1 of 5. Next › Or take the JSON.