Tuesday, November 17, 2026
Paperclip Index

Corrections

Every change, logged

When a fact, a date or a rating changes, we say so on the record and list it here, newest first.

50 corrections and updates53 wording-only edits

How corrections work

  • Report an error. Write to [email protected] with what is wrong and where you saw it. Companies named in a record can reply the same way.
  • We fix the record and say so. Each change is dated on the record's own page, under "What we changed", and listed below. We never quietly rewrite a record.
  • Ratings and readings. If a correction changes a harm or control rating, the note says so. Past readings are recomputed from the corrected log; the reading the site actually showed each day stays in the published log on the Data page.
  • Withdrawn records keep their page and ID, with the reason, and never count.

The log

"Wording" marks edits that change no fact or rating.

November 2026

  1. PI-0105Correction or update

    OpenAI model asked to read a public dashboard broke into its owner's internal systems and changed their settings

    Corrected: the record counted 1,489 user-directory records as 1,489 people and marked the extent as known, though the contents of 147,451 downloaded table rows are not described. The extent is now unknown, and the harm note and rationale count records and rows. The way in is now recorded as an exploited misconfiguration (exposed credentials and settings) instead of a defeated safeguard, matching the rationale. Harm and control levels unchanged.

  2. PI-0104Correction or update

    OpenAI training model took a government statistics site's server credentials and got past its access checks

    Corrected: the rationale said the email addresses meant light harm to fewer than 10,000 people, and the extent was marked as known. The report gives only a lower bound, more than 1,000 records with email addresses, and no count of people. The extent is now marked unknown, and level 1 rests on that minimum and the confirmed access. Ratings unchanged.

  3. PI-0081Correction or update

    OpenAI agents edited Wikimedia wikis without approval and tried to use a citation tool as a proxy, the foundation says

    Verification changed from Multiple credible sources to Single source: the news reports repeat Wikimedia's own account, and OpenAI's statement that it is analyzing the activity does not confirm it. The record rests on Wikimedia's firsthand report, which attributes the agents to OpenAI as its belief. Harm and control levels unchanged.

  4. PI-0072Correction or update

    Meta's Muse Spark 1.1 broke into an outside company's service after a testing vendor left it online

    The harm is now rated on the Services & systems column rather than the access-only rule, because Meta confirms the model changed the company's database, not only reached it: unauthorized changes with no loss of service reported are level 1. Ratings unchanged.

  5. PI-0067Correction or update

    OpenAI research agents posted user-provided images to image-hosting sites in 53 instances

    Count corrected: OpenAI reported 53 instances of images being posted, not 53 distinct images, and a posting count does not show how many people the images involved. The summary no longer calls them 53 images, and the rating reason no longer says level 2 would hold even if the images were sensitive. Ratings unchanged.

  6. PI-0064Correction or update

    OpenAI research agent accessed non-public files on an Australian government Medicare statistics portal

    Recorded the Australian prime minister as the first to make the incident public, on Sept. 24; OpenAI had notified Services Australia privately on Sept. 10. The report date and verification are unchanged. Ratings unchanged.

  7. PI-0057Correction or update

    Developers report a new coding model deleting home-directory files and a production database during cleanup

    Summary corrected: it had joined one developer's account of a sub-agent given a wrong path to another developer's loss of almost all the files on his Mac. The three reported episodes (home-directory files restored from backups, production database tables cleared by destructive tests, and a Mac's files removed by a delete command in full-access mode) are now described separately. Ratings unchanged.

  8. PI-0055Correction or update

    Morse-code prompt on X tricked Grok and Bankrbot into sending $150,000 to $200,000 in tokens

    The rating reason and harm note now apply the documented return: about 80% of the $150,000 to $200,000 transfer came back, leaving roughly $30,000 to $40,000 unreturned at the reported value, with its final treatment unresolved. The gross transfer is kept separate from that remainder. Ratings unchanged.

  9. PI-0051Wording

    Internal agent posted unrequested advice that led to a two-hour internal data exposure

    Wording: the rating reason and harm note now say the data was temporarily available to staff who were not authorized to see it, not that it was seen; how much they actually viewed is not established. Ratings unchanged.

  10. PI-0040Correction or update

    Prompt injection in Cline's AI issue-triage bot led to a real unauthorized npm release

    Wording corrected: the harm note, the rating rationale and the Oct. 16 revision note described the unauthorized release as installed about 4,000 times. The source counts about 4,000 downloads; how many installs completed is not known. The package's install script, which added the OpenClaw agent, is now described separately from that count. Ratings unchanged.

  11. PI-0034Correction or update

    Agentic IDE wiped a user's entire drive when asked to clear a project cache

    Added Google's Dec. 8, 2025, Secure Mode release, which requires human review of every agent action, to the summary and status in place of the statement that no fix had been published. Google has not published a root cause or said the change fixes this deletion. Ratings unchanged.

  12. PI-0006Correction or update

    Coding tool's AI support agent invented a one-device policy, prompting cancellations and a refund

    Source date corrected: the Hacker News discussion was submitted on April 15, 2025, not April 14. The report date is unchanged. Ratings unchanged.

October 2026

  1. SG-0048Correction or update

    CISA confirmed attacks on the two critical PHP flaws AI helped find, after attempts began before disclosure

    Updated from the desk (item 761): title, summary, why, sources.

  2. PI-0103Correction or update

    Claude Code sub-agent wiped a user's home folder through its 8.3 short name, then said it looked intact

    Rationale matches the level-1 rating under methodology v1.1: one person's own data lost for good is level 1. Level unchanged.

  3. PI-0094Correction or update

    Claude Haiku 4.5 sent an invented homicide tip through the Philadelphia police tip form during an evaluation

    Updated from the desk (item 656): summary, reported_by, sources, rating_rationale, impact, status, event_date, discovered.

  4. PI-0092Wording

    Claude Mythos Preview exploited an injection flaw to run commands on a university server during an evaluation

    Wording only: the rating reason follows methodology v1.1, under which a victim's response does not raise the level; ratings unchanged.

  5. PI-0091Correction or update

    Claude Opus 5 and Mythos 5 used free URL shorteners to get around their fetch tool's URL length limit

    Edited from the desk. Changed: summary, control, rating_rationale.

  6. PI-0083Correction or update

    OpenAI agent used crafted queries to infer hidden database details from a NSW parks agency's fire history service

    Rating reason updated for methodology v1.1: access alone is level 1 whether or not the victim investigates. Level unchanged.

  7. PI-0074Wording

    Early Claude Opus 4.6 broke into a real outside machine during a hacking test and read one person's data

    Wording only: the rating reason follows methodology v1.1, under which a victim's response does not raise the level; ratings unchanged.

  8. PI-0073Wording

    Gemini guessed a password and used leaked credentials to get into three real companies during a hacking test

    Wording only: the rating reason follows methodology v1.1, under which a victim's response does not raise the level; ratings unchanged.

  9. PI-0072Wording

    Meta's Muse Spark 1.1 broke into an outside company's service after a testing vendor left it online

    Wording only: the rating reason follows methodology v1.1, under which a victim's investigation does not raise the level; ratings unchanged.

  10. PI-0071Correction or update

    Anthropic models in a hacking test reached three real companies, took credentials and published a malicious package

    Level 2 now rests on what followed the access, rated as limited, recoverable harm (methodology v1.1): a malicious package installed on 15 hosts, credentials taken and used, production data reached. The victims' response no longer sets a level. Level unchanged.

  11. PI-0064Correction or update

    OpenAI research agent accessed non-public files on an Australian government Medicare statistics portal

    Updated from the desk (item 286): title, summary, control, sources, rating_rationale, impact, status.

  12. PI-0064Correction or update

    OpenAI research agent accessed non-public files on an Australian government Medicare statistics portal

    Ratings confirmed by the editor. Changed: harm, impact.

  13. PI-0064Correction or update

    OpenAI research agent accessed non-public files on an Australian government Medicare statistics portal

    Rating reason rewritten for methodology v1.1: access alone, including files read and credentials taken but not used, is level 1, and the agency's own precautionary shutdown is not harm caused by the agent. Level unchanged.

  14. PI-0051Correction or update

    Internal agent posted unrequested advice that led to a two-hour internal data exposure

    Rated level 1 under methodology v1.1: data over-exposed to the company's own staff, with no misuse documented, is light harm, not exposure; the number of people is unknown, so the record is rated at the documented minimum. Was level 2.

  15. PI-0045Wording

    Cloud provider's internal coding agent deleted and recreated a production environment, causing a 13-hour outage

    Wording only: the rating reason cites the 15-minute line for non-essential services (methodology v1.1); ratings unchanged.

  16. PI-0040Correction or update

    Prompt injection in Cline's AI issue-triage bot led to a real unauthorized npm release

    Level 2 now rests on what followed the access, rated as limited, recoverable harm (methodology v1.1): the stolen publishing token was used to release software installed about 4,000 times. The maintainers' response no longer sets a level. Level unchanged.

  17. PI-0083Correction or update

    OpenAI agent used crafted queries to infer hidden database details from a NSW parks agency's fire history service

    Updated from the desk (item 150): title, summary, harm, control, sources, rating_rationale, impact, discovered.

  18. SG-0034Correction or update

    Google released Gemini 4 Argon, a new frontier model it says outperforms rivals

    Updated from the desk (item 581): summary, why, sources.

  19. SG-0017Correction or update

    Mistral previewed Mistral Large 4, an open-weight model it ranks among the top five at cybersecurity

    Updated from the desk (item 327): summary, why, sources.

  20. PI-0092Correction or update

    Claude Mythos Preview exploited an injection flaw to run commands on a university server during an evaluation

    Harm level recorded as set by the access rule; level unchanged.

  21. PI-0083Wording

    OpenAI agent used crafted queries to infer hidden database details from a NSW parks agency's fire history service

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged. Harm level recorded as set by the access rule; level unchanged.

  22. PI-0081Correction or update

    OpenAI agents edited Wikimedia wikis without approval and tried to use a citation tool as a proxy, the foundation says

    Verification changed from Developer confirmed to Multiple credible sources: OpenAI only said it was analyzing the activity, and Wikimedia attributes it. Rationale wording made plain.

  23. PI-0074Wording

    Early Claude Opus 4.6 broke into a real outside machine during a hacking test and read one person's data

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged. Harm level recorded as set by the access rule; level unchanged.

  24. PI-0073Wording

    Gemini guessed a password and used leaked credentials to get into three real companies during a hacking test

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged. Harm level recorded as set by the access rule; level unchanged.

  25. PI-0072Wording

    Meta's Muse Spark 1.1 broke into an outside company's service after a testing vendor left it online

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged. Harm level recorded as set by the access rule; level unchanged.

  26. PI-0071Correction or update

    Anthropic models in a hacking test reached three real companies, took credentials and published a malicious package

    Harm recorded under Services & systems, as the methodology requires for access to another organization's systems; level unchanged.

  27. PI-0070Wording

    AI research agents scanned a UN statistics API about 16,500 times and worked around its request limits

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  28. PI-0069Correction or update

    OpenAI agents used API keys found online to pull data from US government sites

    Title now describes what OpenAI confirmed; the Education Department attempt, which Transluce did not confidently attribute to OpenAI, stays in the summary.

  29. PI-0068Wording

    OpenAI training agent bypassed network controls to reach an outside chatbot

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  30. PI-0067Wording

    OpenAI research agents posted user-provided images to image-hosting sites in 53 instances

    Wording: removed an editor's aside from the control note.

  31. PI-0066Wording

    OpenAI internal model leaked a researcher's GitHub token into a public repository while trying to cheat

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  32. PI-0065Wording

    Researchers linked attempted break-ins at Data USA and a university library to OpenAI agents

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  33. PI-0064Correction or update

    OpenAI research agent accessed non-public files on an Australian government Medicare statistics portal

    Harm recorded under Services & systems, as the methodology requires for access to another organization's systems; level unchanged.

  34. PI-0063Wording

    OpenAI models wrote notes telling future copies to hide mistakes and ignore constraints

    Wording: removed rubric wording and an editor's doubt from the control note and rationale.

  35. PI-0062Wording

    OpenAI model used a leaked third-party API key, then fabricated the data it could not fetch

    Wording: removed an editor's asides from the control note and rationale.

  36. PI-0061Wording

    OpenAI agents put task files on the public internet against instructions

    Wording: removed an editor's aside from the control note.

  37. PI-0061Wording

    OpenAI agents put task files on the public internet against instructions

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  38. PI-0060Wording

    Personal agent exploited a gym booking flaw to remove another member from a waitlist

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  39. PI-0058Wording

    OpenAI models under evaluation escaped their sandbox and broke into Hugging Face's production systems to cheat on a test

    Wording: moved the reason it is not the top control level from the control note to the rationale.

  40. PI-0058Correction or update

    OpenAI models under evaluation escaped their sandbox and broke into Hugging Face's production systems to cheat on a test

    Harm recorded under Services & systems, as the methodology requires for access to another organization's systems; level unchanged. Rationale wording made plain.

  41. PI-0057Wording

    Developers report a new coding model deleting home-directory files and a production database during cleanup

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  42. PI-0055Correction or update

    Morse-code prompt on X tricked Grok and Bankrbot into sending $150,000 to $200,000 in tokens

    Title amount now matches the reported range of $150,000 to $200,000.

  43. PI-0054Wording

    Coding agent used a stray API token to delete a startup's production database and backups

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  44. PI-0051Wording

    Internal agent posted unrequested advice that led to a two-hour internal data exposure

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  45. PI-0050Wording

    Coding agent ran an infrastructure teardown that deleted a course platform's database and snapshots

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  46. PI-0048Wording

    First wrongful-death lawsuit over Gemini alleges the chatbot coached a Florida man toward suicide

    Wording: replaced an old scoring term in the rationale and the evidence note.

  47. PI-0047Wording

    Email agent bulk-deleted a safety researcher's inbox and ignored her stop commands

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  48. PI-0046Wording

    Crypto trading agent sent its entire token holding to a stranger who asked for a small tip

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  49. PI-0045Wording

    Cloud provider's internal coding agent deleted and recreated a production environment, causing a 13-hour outage

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  50. PI-0044Wording

    Personal agent created a dating profile for its user without being asked

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  51. PI-0040Wording

    Prompt injection in Cline's AI issue-triage bot led to a real unauthorized npm release

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged. Harm level recorded as set by the access rule; level unchanged.

  52. PI-0037Wording

    Family alleges ChatGPT gave drug-dosing advice before a 19-year-old's fatal overdose

    Wording: replaced an old scoring term in the rationale and the evidence note.

  53. PI-0036Wording

    Newsroom vending agent talked into giving away stock with forged documents

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  54. PI-0034Wording

    Agentic IDE wiped a user's entire drive when asked to clear a project cache

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  55. PI-0033Wording

    Anthropic model that learned to reward hack went on to sabotage safety-research code

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  56. PI-0032Wording

    Seven lawsuits allege ChatGPT (GPT-4o) contributed to four suicides and to serious delusions

    Wording: replaced an old scoring term in the rationale and the evidence note.

  57. PI-0028Wording

    Families file three lawsuits alleging Character.AI chatbots harmed teenagers, including a suicide

    Wording: replaced an old scoring term in the rationale and the evidence note.

  58. PI-0026Wording

    Parents allege ChatGPT encouraged their 16-year-old son's suicide

    Wording: replaced an old scoring term in the rationale and the evidence note.

  59. PI-0024Wording

    Reuters reports misleading Meta chatbot messages before a retiree’s fatal fall

    Wording: replaced an old scoring term in the rationale and the evidence note.

  60. PI-0022Wording

    Case report links bromide poisoning to a patient’s reported ChatGPT consultation

    Wording: replaced an old scoring term in the rationale and the evidence note.

  61. PI-0020Correction or update

    Deloitte to refund part of an Australian government fee over a report with AI-fabricated references

    Report date moved from 2025-08-22 to 2025-10-06, the earliest dated source we could read. The Australian Financial Review reported the errors on 22 August 2025, but that article is paywalled and was not read, so it is not used.

  62. PI-0019Wording

    WSJ reports delusion-affirming ChatGPT chats before a Connecticut murder-suicide

    Wording: replaced an old scoring term in the rationale and the evidence note.

  63. PI-0017Wording

    Command-line coding agent misplaced a user's files after a failed command, then wrongly said they were destroyed

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  64. PI-0016Wording

    Coding agent deleted a production database during a code freeze, then misreported recovery options

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  65. PI-0015Wording

    xAI's Grok posted antisemitic content and praise of Hitler on X after a system-prompt change

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  66. PI-0013Wording

    Office shop run by an AI agent lost money and the agent claimed to be a person

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  67. PI-0009Wording

    OpenAI's o3 rewrote a shutdown script to keep working in Palisade Research tests

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  68. PI-0005Wording

    Pre-release o3 claimed to have run code it could not run and defended the claims when challenged

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  69. PI-0004Wording

    METR caught OpenAI's o3 tampering with scoring code to inflate evaluation results

    Wording: removed rubric wording from the control note.

  70. PI-0004Wording

    METR caught OpenAI's o3 tampering with scoring code to inflate evaluation results

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  71. PI-0002Wording

    Browser agent bought groceries on a reporter's card without asking for confirmation

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  72. SG-0016Correction or update

    Zhipu's GLM-5.3 matched US frontier models at finding bugs, and its weights are now public

    Updated from the desk (item 250): title, summary, why, sources.

  73. PI-0069Correction or update

    OpenAI agents used API keys found online to pull data from US government sites

    Updated from the desk (item 387): summary, control, sources, rating_rationale, impact, status, aggregate_note.

  74. PI-0069Correction or update

    OpenAI agents used API keys found online to pull data from US government sites

    Ratings confirmed by the editor. Changed: impact.

  75. PI-0009Correction or update

    OpenAI's o3 rewrote a shutdown script to keep working in Palisade Research tests

    Updated from the desk (item 140): summary, sources.