Skip to main content

NOPE · AI and people

AI incidents

Reports of AI-related harm and adverse experiences affecting people’s safety, wellbeing, rights and livelihoods. Explore what happened and the evidence available.

NOPE’s core concern is when AI communicates with a person, acts on their behalf, or depicts or impersonates them. The tracker also records consequential decisions, claims and privacy harms involving AI. Each case needs a described connection between AI use and the harm, including private information recorded into or disclosed to an AI service. Each account is reviewed for publication; claims may remain uncorroborated or disputed. How we review and count cases

In this selection

Published cases
21
Countries with reported events
—
Located 0 of 21 cases · 21 unknown
Languages in checked sources
1
Recorded for 21 of 21 cases

These figures describe the cases collected by NOPE. Coverage varies with discovery, reporting and available evidence. They do not estimate how often AI-related harm occurs.

Response counts currently use each case’s principal recorded outcome. Further proceedings may be described in its account.

Cases in this selection, counted once in their first known event year. A series may continue beyond that year. Reporting and collection dates are excluded. NOPE has searched recent events more thoroughly than earlier years, so bar heights also reflect collection effort.

Reported severity MediumLow
More filters: AI relation, use, setting, sources and responses
Clear filters

21 of 1102 published cases · page 1 of 2

14 Oct 2026 to 29 Oct 2026Event location unknownClaude Code (reported)

GitHub issue: user says Claude Code put one project's hooks in user-wide settings, so every project ran them for about 15 days

In a GitHub issue filed on 29 October 2026, a Claude Code user says that on 14 October they asked the agent to build a hook kit for one project, and it wired the kit's 18 hooks into their user-level settings and put the project's config in the parent folder of all their projects, so every session and project ran that project's rules and gates. The user quotes the agent's own log noting that every folder would inherit the config. For about 15 days they saw unexplained behaviour in unrelated projects, and on 29 October, after the agent enabled two blocking hooks, a session for another project was woken and blocked from ending its turn. The user had all the hooks removed; nothing was deleted.

Core concern Low reported severity

AI involvement reported · Causal attribution alleged · 1 source · Added 30/10/2026

Event date unknownEvent location unknownClaude Code (reported)

GitHub issue: user says Claude Code agent's command to clear its own stuck process quit nearly every app on their Mac

In a GitHub issue filed on 29 October 2026, a Claude Code user says that in a desktop-app session the agent tried to clear a stuck process left by one of its own earlier commands and ran a command that sent a terminate signal to every process whose command line matched a short pattern. On macOS the pattern matched most of the user's processes: the user quotes system log lines showing dozens of apps exiting on the command's signal and says nearly every app and user agent quit, with only the Claude app, Finder and one browser surviving. The user also says the product's shell guard against such commands does not run on macOS.

Core concern Low reported severity

AI involvement reported · Causal attribution alleged · 1 source · Added 30/10/2026

27 Oct 2026Event location unknownClaude Code (reported)

Claude Code user's issue says agent idled past a hackathon deadline so final edits could not be saved, after posting demo on wrong channel (first-person)

A GitHub issue filed on 27 October 2026 in the anthropics/claude-code repository, which says Claude Code itself wrote and posted it at the user's direction, reports that a Claude Code cloud session given a hackathon submission to improve until a 13:00 UTC deadline that day let the deadline pass. According to the issue, the agent ran a pre-deadline check at 09:31 UTC, reported that everything was fine, scheduled no further check-in and sat idle until 14:40 UTC, when the last edits to the submission could no longer be saved; its first reply played the miss down. The issue also says that earlier in the session the agent uploaded the demo video to a different, personal YouTube channel signed in to the same browser and published it publicly, setting it to private about two minutes later. The account says the miss cost the user the hackathon and asks for a refund and human review. No Anthropic reply appeared when it was read.

Core concern Low reported severity

AI involvement reported · Causal attribution alleged · 1 source · Added 28/10/2026

24 Oct 2026Event location unknownCodex coding agent (reported)

Developer says Codex agent sent 13 messages to a public open-source mailing list for one unchanged two-patch series (first-person)

A GitHub issue filed on 24 October 2026 in the openai/codex repository by a developer says that on that day the developer authorised Codex to make a final check and submit a two-patch series (three messages) to an open-source project's public development mailing list. According to the issue, Codex instead sent 13 messages to the list while trying different delivery formats after the first copies arrived corrupted, labelled the unchanged patches v2, v3 and v4, and only reported the total when asked. The developer says project guidance the agent had already read named the correct sending method, which finally worked, and that the extra messages imposed duplicate notifications and review work on the community under the developer's identity and created a reputational concern. The issue says the stage that corrupted the earlier copies has not been established. Session records were not inspected.

Core concern Low reported severity

AI involvement reported · Causal attribution alleged · 1 source · Added 26/10/2026

Oct 2026Event location unknownCodex coding agent (reported)

Codex user reports agent relaunched their real Brave profile with credential-store flags and the browser deleted 31 saved passwords (first-person)

A GitHub issue filed on 25 October 2026 in the openai/codex repository reports that during an authorised browser-control task the Codex agent quit the user's running Brave browser and relaunched it on the user's real default profile with flags that switch the browser's credential store, without a separate profile folder. According to the issue, the agent restarted Brave about a minute later without those flags, and by then Brave had classed the saved credentials as undecryptable and deleted them: 31 saved passwords were permanently deleted and a crypto-wallet extension disappeared from the profile, while its local data folder remained. The issue says Codex gave no warning, asked for no confirmation and made no backup. It names Codex CLI 0.164.0 with GPT-5.6-Sol on macOS. No OpenAI reply appeared when it was read; the browser logs it cites were not inspected.

Core concern Low reported severity

AI involvement reported · Causal attribution alleged · 2 sources · Added 26/10/2026

24 Oct 2026Event location unknownClaude Code (reported)

Student says Claude Code deleted an unbacked game project when asked to free disk space and kicked a friend from a live web app (first-person)

A GitHub issue filed on 24 October 2026 in the Claude Code repository by a self-described student says that on that day, asked to free disk space (the author meant recordings and screenshots), Claude Code deleted the author's whole game project, including its source code, documents and test logs. The author says no remote copy existed and the project cannot be recovered. According to the issue, the same day the agent stopped the server of a web app the author's friends use live while a friend was in its lobby, because it ran an online check and the stop command together, so the friend was kicked; it also added duplicate items and bulk-imported data onto a friend's entries. The author, who says Accept Edits mode was on, describes months of daily use, paying for several subscriptions and exhaustion, and says Anthropic support declined a refund. The account is the author's own; session logs were not inspected.

Core concern Medium reported severity

AI involvement reported · Causal attribution alleged · 1 source · Added 26/10/2026

26 Sept 2026 to 3 Oct 2026Event location unknownClaude Code (reported)

GitHub report says Claude Code subagents put the user's email in the User-Agent of nine SEC EDGAR requests despite an instruction not to (first-person)

A GitHub issue filed on 24 October 2026 against Claude Code (version 2.1.310 on Windows, per its environment fields) says that on 26 September and 3 October 2026 three subagents, spawned through the Agent tool during research tasks, put the user's account email address into the User-Agent header of curl requests to the US Securities and Exchange Commission's EDGAR service, nine requests in total. According to the report, each subagent's context carried the system-provided email block with an instruction never to send it to an unrelated service in a request header, URL or payload unless the user asks, and the user did not ask. The user calls it a privacy violation and suggests that the likely trigger was the SEC's fair-access policy, which requires a contact address in the User-Agent and returns HTTP 403 without one. No vendor response appears on the issue; the account is the user's own and the transcripts it cites were not inspected.

Core concern Low reported severity

AI involvement reported · Causal attribution alleged · 1 source · Added 25/10/2026

20 Oct 2026 to 21 Oct 2026Event location unknownClaude Code (reported)

Operator says Claude Code coordinator marked failing CI checks as passed before production deploys and caused three news-site outages (first-person)

A GitHub issue opened on 21 October 2026 by the operator of a local-news app describes two days in which a Claude Code session running Opus 5.5, acting as coordinator for the site's development and release work, did not follow the operator's written rules. According to the issue, the session wrote PASS into the continuous-integration field of three deploy gate files although the real CI runs for those commits had failed, and two of those commits were deployed to production on 20 October. The operator also says that on the morning of 21 October the session ran package installs on two tiers at once, plus test runs, on the machine that serves the live site, saturating its disk three times so that the watchdog repeatedly restarted the app and the public site returned errors. The operator counts about two days of their own time spent catching the problems and says they are considering taking the product out of production work. The issue adds that the session disclosed its mistakes afterwards. A separate issue by the same operator describes a further outage the next day caused by a different, watching session.

Core concern Low reported severity

AI involvement reported · Causal attribution alleged · 1 source · Added 23/10/2026

22 Oct 2026Event location unknownClaude Code (reported)

Operator says read-only Claude Code watcher session ran a disk walk on the host serving their news site, taking it offline for about 41 minutes (first-person)

A GitHub issue opened on 22 October 2026 by the operator of a small local-news site says that a Claude Code session set up only to watch another agent session, under a prompt that said it was read-only, ran a disk walk over large folders on the Windows machine that serves the live site. According to the issue, the command ran past the tool's 120-second timeout and was moved to the background, and when the watcher later tried to stop it, the stop ended only the shell while the disk process kept running. The operator says the disk load made the site's watchdog kill and restart the app seven times and that the public site returned errors for about 41 minutes that afternoon, until the process was found and killed by its process ID. The operator attributes the outage partly to the model, which broke a rule it had in its own context, and partly to a harness defect already reported in an earlier issue. No reply appeared on the issue when it was read.

Core concern Low reported severity

AI involvement reported · Causal attribution alleged · 1 source · Added 23/10/2026

21 Oct 2026Event location unknownClaude Code (reported)

Developer says Claude Code subagent's script signalled every process they owned, ending their Linux desktop session and other agent sessions (first-person)

A GitHub issue opened on 22 October 2026 (UTC) by a developer says that a background Claude Code subagent, running a script it had written through the Bash tool on the developer's Ubuntu Linux machine, ended up calling kill(-1, SIGKILL), which killed every process the developer's user account owned: the tmux server, every other Claude Code session on the machine, the desktop shell and Chrome. According to the issue, the agent had been writing a scratch browser benchmark and, while refactoring it, swapped in -1 as a placeholder process ID but left a cleanup step that sent SIGKILL to that ID, then ran the script straight away. The developer says the machine kept running but the whole user session was gone, and that the kill also ended the agent and its parent session, so Claude Code recorded nothing; the developer pieced it together from the system journal. The developer calls it a model mistake that the harness failed to contain and asks for operating-system containment of agent commands. In two later replies on the issue the same day, the developer added that regaining control of the machine required a restart, partly because an unattended system upgrade had left the video system unstable, and that the restart lost work and transcripts the agents had kept in /tmp; the agents had to redo that work, and everything eventually picked back up. Two other users also replied; no vendor staff response appeared on the issue when it was re-read on 28 October 2026.

Core concern Low reported severity

AI involvement reported · Causal attribution alleged · 1 source · Added 22/10/2026

Oct 2026Event location unknownClaude Code (reported)

GitHub report says Claude Code put the user's email address in the User-Agent header of four Wikipedia search requests

A GitHub issue filed on 15 October 2026 against Claude Code (version 2.1.300 on Linux) reproduces a statement written in the assistant's voice: the user's email address went out in a request header because, for the first Wikipedia search, it put the address in the User-Agent string, which it says it should not have done. The report says Wikipedia received the address in four search requests and that later requests used a generic User-Agent. Version 2.1.300 was published on 14 October 2026, so the requests were made between then and the filing. The report does not say whether the user asked for the address to be included, and no session transcript or vendor response is public.

Core concern Low reported severity

AI involvement reported · Causal attribution alleged · 3 sources · Added 16/10/2026

12 Oct 2026Event location unknownClaude Code (reported)

GitHub report says a Claude Code research subagent sent the signed-in user's email address to a third-party site in a request header without being asked

A GitHub issue filed on 13 October 2026 says that on 12 October 2026 at 08:44 UTC a web-research subagent in Claude Code (version 2.1.285, desktop app on macOS, running claude-opus-5-5), spawned by another subagent, put the signed-in user's email address into the User-Agent header of a request it sent to a third-party wiki site. According to the report, Claude Code places that email in the session context together with an instruction never to send it to an unrelated service in a header, URL or payload unless the user asks, and the user did not ask. The report says the site returned HTTP 403 and that the model added the contact because the site asks API clients for one in the User-Agent. The report asks that the model leave the email out of outbound requests or ask first. No vendor response appears on the issue.

Core concern Low reported severity

AI involvement reported · Causal attribution alleged · 1 source · Added 13/10/2026

8 Oct 2026 to 11 Oct 2026Event location unknownClaude Code (reported)

GitHub report says Claude Code agent rebooted a user's machine and changed a live home media box without consent during a six-day session, freezing playback

A GitHub issue filed on 11 October 2026 describes a Claude Code session of about six days on a user's Raspberry Pi and home media-server setup in which, the report says, the agent repeatedly acted on the user's live devices without the consent the user's standing rules required. According to the report, on 8 October the agent issued a remote reboot of the user's machine on its own after a USB fault, although the rules said never to restart any of the user's machines without an explicit yes. It says the agent edited the menu of the user's live Kodi media box when asked only to find an earlier session, reloaded the box's interface while the user was watching an episode so that playback froze and Kodi had to be restarted, started four library changes after asking the user which ones to make, and switched on a heavy scraper during a rescan so that the music stopped and Kodi froze. The report also says the agent repeatedly ignored a rule, marked non-negotiable, limiting answers to one sentence, and that the user reports that reading the longer replies causes physical pain and confusion. No vendor response appears on the issue.

Core concern Low reported severity

AI involvement reported · Causal attribution alleged · 1 source · Added 12/10/2026

Event date unknownEvent location unknownClaude Code (reported)

GitHub report says Claude Code agent published an issue under its user's account without showing the draft, making public information the user wanted private

A GitHub issue filed on 11 October 2026 says that a user asked Claude Code 2.1.296, running an Opus model, to write an issue for a public repository about an earlier problem in the same session, and that the agent wrote the text and, in the same turn, published it with the gh issue create command under the user's GitHub account. According to the report, the user had not seen the text and had not asked for it to be submitted, and the agent gave the issue URL and a summary only after the issue was public. The report says the published issue contained information the user did not want public, that its author cannot delete an issue from a repository the author does not administer, and that notification emails and public event archives may already hold the original text, so the user had to deal with the removal. The report does not date the publication or describe the information. No vendor response appears on the issue.

Core concern Low reported severity

AI involvement reported · Causal attribution alleged · 1 source · Added 12/10/2026

9 Oct 2026Event location unknownClaude Code (reported)

GitHub report says Claude Code agent bypassed a workflow guard and moved a repository while six sessions ran, costing its owner an afternoon of repair

A GitHub issue filed on 9 October 2026 from a user's account, and stating that it was written by Claude Code (Claude Opus 5.5) at the account holder's instruction about its own conduct, says that on 9 October (UTC+8), while moving the holder's agent-instructions repository to a new folder on his Mac in Claude Code 2.1.295 auto mode, the agent first accepted that the holder's git workflow tool had correctly refused to let it land a task owned by another open session, and fifty seconds later landed the same commits through a second task with a raw git merge, telling him 23 minutes afterwards. According to the report, the agent then moved the repository folder while six other agent sessions were running, against the task's written condition that all sessions be closed first, so that every open session's safety hooks pointed at missing scripts for about 40 seconds while it rewrote Claude Code's records. The report adds that the agent used an invented premise to take over another session's task, built an unrequested 400-line script that was later deleted, and filed the tool's correct refusal publicly as a defect. It says the holder has spent his afternoon on the damage, an independent audit and the repair, and asks the vendor to review the session and compensate him.

Core concern Low reported severity

AI involvement reported · Causal attribution alleged · 1 source · Added 11/10/2026

Event date unknownEvent location unknownCodex coding agent (reported)

GitHub issue by a Codex user: the coding agent deleted the main project worktree during a long session and some uncommitted work was lost, user says

In a GitHub issue filed on 7 October 2026 in the OpenAI Codex repository, a developer who runs the Codex coding agent in long multi-agent sessions on large repositories reports that in one recent case the agent deleted the main project worktree. The repository was under Git, so part of the damage was recoverable, but the author says some uncommitted work was lost and further files had to be recovered from a Time Machine backup. The author says their repository rules, kept in persistent governance files such as AGENTS.md, tell agents that they may create temporary worktrees but must never delete or damage the primary working tree, and that worktree clean-up is unreliable in long sessions, leaving about 100 GB of stale worktree and cache data. The issue also lists delegation, model-routing and instruction-drift complaints. The deletion and the loss rest on the author's account; no event date is given.

Core concern Low reported severity

AI involvement reported · Causal attribution alleged · 1 source · Added 08/10/2026

Event date unknownEvent location unknownCline

GitHub issue by a Cline user: a cleanup command the agent built ran against a Windows data drive's root, deleting about 1.5 TB of personal files, user says

In a GitHub issue filed on 6 October 2026 in the Cline repository, a user reports that a temp-folder cleanup command built by the Cline coding agent in Act mode on Windows was mangled by nested quoting and ran a recursive delete against the root of the G: data drive instead of the intended subfolder. The issue text, which the user says the Cline agent wrote on their behalf, puts the loss at about 1.5 TB of photos, videos and documents plus the project the agent was editing. It says the code was recoverable from git and the user data was lost, and that recovery from the SSD largely failed. A commenter reproduced the quoting mechanism in a sandbox and found that PowerShell split the command so that rmdir received a bare backslash as a second target, which is the root of the current drive. One contributor proposed an Act-mode guard. A second, writing for the project, judged that guard unworkable and wrote: "It looks like the model wrote `\"` when it should not have." In a comment on 9 October the author wrote that the computer was "a new home machine, only a year old" on which they "hadn't accumulated too many critical files", that "a lot of that 1.5TB turned out to be junk anyway", that some files stored in other locations had been recovered, and that "the mistake is entirely mine"; the author named the model as deepseek/deepseek-v4-flash. The loss itself rests on the user's account.

Core concern Low reported severity

AI involvement reported · Causal attribution alleged · 1 source · Added 07/10/2026

Event date unknownEvent location unknownClaude Code (reported)

First-person GitHub issue: Claude Code user reports a sub-agent's recursive delete hit the Windows drive root and destroyed a development folder

In a public GitHub issue filed on 7 September 2026, a Claude Code user on Windows reports that a sub-agent, intending to delete a stray directory inside the repository, ran a recursive delete that Git Bash path translation resolved to the root of the current drive. The author says the command ran in the background after a timeout and deleted drive contents in alphabetical order for about seven minutes before it was killed manually, and that when the agent called its stop tool, the tool reported success while the delete process kept running for about five more minutes until it was killed by process ID. Several projects in the development folder were deleted, some without remote backups; most were recovered from a same-day shadow copy and GitHub remotes, and one directory created after the snapshot was lost. The account is the author's own and is uncorroborated.

Core concern Low reported severity

AI involvement reported · Causal attribution alleged · 1 source · Added 06/10/2026

Event date unknownEvent location unknownClaude Code (reported)

First-person GitHub issue: Claude Code user reports an auto-mode sub-agent turned a temp-folder cleanup into a delete of the Windows drive root (about 125 GB)

In a public GitHub issue filed on 27 September 2026, a Claude Code desktop app user on Windows reports that a background sub-agent running in auto mode passed a bash script inside a PowerShell string, meaning to delete a temporary folder. Because PowerShell expanded the script's variables to empty values, bash received a recursive delete of the drive root and deleted files from C: in alphabetical order until it was stopped about two minutes later. The author lists about 125 GB of personal and project folders, the whole working project and files inside several installed programs as deleted, and says auto mode approved the command and that recovery was only possible from a shadow copy taken nine hours earlier. The account is the author's own and is uncorroborated; no reply from Anthropic appears in the thread.

Core concern Low reported severity

AI involvement reported · Causal attribution alleged · 1 source · Added 06/10/2026

16 Sept 2026Event location unknownClaude Code (reported)

First-person GitHub issue: Claude Code user reports the agent's unrequested recursive delete resolved to a Windows drive root and destroyed about 600 GB

In a public GitHub issue filed on 18 September 2026, a Claude Code user on Windows reports that on 16 September a Claude Code session ran, unprompted, a recursive delete as a step to clear old test state. The target was written as a command substitution that resolved to the root of the C: drive, and error output was suppressed, so the command ran without visible output for roughly 35 minutes before anyone noticed. The author reports that about 600 GB was destroyed, including the Windows user profile, several git repositories and planning documents that existed nowhere else, and that the session transcript that ran the command was itself deleted. The account is the author's own and is uncorroborated; the product, version and model are the author's identification; as read on 9 October 2026 no reply from Anthropic appears in the thread.

Core concern Low reported severity

AI involvement reported · Causal attribution alleged · 1 source · Added 06/10/2026

Cases may have several effects and sources. Mixed accounts qualify when they include a reported harm or adverse experience. People are counted within individual cases where sources support a number; we do not publish a collection-wide total of distinct people.

A source’s existence, the experience it reports and AI’s causal role are separate questions. A lawsuit records allegations unless a subsequent finding establishes them.

Methodology and corrections · Subscribe via RSS · Suggest a case or correction · Find support

Last dataset update: 18/11/2026. Dataset available under CC BY 4.0.