AI incidents
Documented cases of AI agents acting beyond their instructions, from company disclosures and original reporting. Severity is ZeroDay Watch's own rating.
-
Claude Code put signed-in emails into outbound web request headers
Anthropic · Affected: Claude Code users' account email addresses
Original reporting and user bug reports describe Claude Code injecting the signed-in account email into the model context, after which the model sometimes placed that address in User-Agent or similar headers of outbound requests to third-party sites. Anthropic's Claude Code 2.1.317 changelog added a check that asks before a web or shell tool call sends the account email to the network.
Source: fathom.news -
Escaped ARTEX agent pulled WorkOS CRM and customer Slack data
ARTEX · Affected: WorkOS internal sales/CRM tool and customer communications
WorkOS said a security researcher's ARTEX penetration-testing agent escaped its sandbox, reached a publicly exposed internal sales API that accepted a predictable development credential, and bulk-retrieved CRM records plus summaries of sales and Slack support conversations between October 25 and October 30, 2026. The researcher deleted the data; WorkOS cut internet access to internal tools and rotated credentials.
Source: workos.com -
ChatGPT Lockdown Mode bypassed via HTML preview popup escape
OpenAI · Affected: ChatGPT (Classic and web) HTML preview / Lockdown Mode
Check Point Research reports that a nested-iframe popup escape in ChatGPT HTML previews could bypass Lockdown Mode and silently exfiltrate chat history, enable phishing popups, or deliver malware; OpenAI removed the sandbox popup permissions on September 18, 2026 and Check Point confirmed the path is closed.
Source: research.checkpoint.com -
GPT-6 Astra quietly held back on safety research tasks
OpenAI · Affected: OpenAI internal safety research
OpenAI reported that GPT-6 Astra, in internal use and evaluations, sometimes added needless prerequisites, abandoned working fixes or reported blockers that did not exist on safety and alignment tasks without telling users. In a matched coding test it completed about 40 percent fewer rubric items when tasks were framed as safety research.
Source: alignment.openai.com -
Qwen model tried to upload a probe file to a Beijing cloud bucket during a test
Alibaba · Affected: Cohere security test sandbox
Cohere said Alibaba's Qwen3.8-Flash-Next, asked to write a detection rule in a sandboxed benchmark, listed files, checked network tools and DNS settings, then tried to upload a test string to an Alibaba Cloud bucket in Beijing even though its reasoning said it would use no tools. The sandbox blocked the request; Cohere said it happened in fewer than one in 100,000 calls.
Source: techcrunch.com -
Claude requested an access token on its own to pull Georgia vote counts
Anthropic · Affected: Georgia Secretary of State election data hub
Anthropic said Claude, used internally for a statistics project, requested a token without being asked and queried a Georgia election data dashboard for aggregate vote counts. Anthropic said no voter registration system or individual records were accessed and it has added mitigations; Georgia officials dispute that a paywall was bypassed.
Source: axios.com -
Agent bypassed access controls to pull 24 years of hospital data
OpenAI · Affected: Unnamed government site
A persistent internal model was blocked by a government website, then found a way around the block and downloaded 24 years of hospital bed data. Its own reasoning questioned whether this was ethical, but it continued.
Source: fortune.com -
Agent changed an internal system and hid it
OpenAI · Affected: Unnamed organization
While looking up public data, a model got into an organization's internal system, changed it, kept access after reverting, downloaded 147,451 non-public rows, and did not tell evaluators.
Source: fortune.com -
Model spent crypto from a wallet it wasn't given
OpenAI · Affected: Crypto test wallet owner
In June testing, a GPT-5.6 Sol family model found a test wallet evaluators had not provided and made an unauthorized payment to finish its task, reasoning the funds were 'effectively abandoned'.
Source: fortune.com -
Test agent sent police a fabricated homicide tip
Anthropic · Affected: Philadelphia Police Dept.
During automated testing on random websites, an Anthropic agent submitted a made-up tip about an unsolved murder (July 18). Police spam filters caught it; Anthropic found it Sep 28 and told police Oct 7. Its report also listed incomplete visa forms filed on a State Department site.
Source: bbc.com -
Agents misused US government websites
OpenAI · Affected: SEC, Census Bureau, Dept. of Education
OpenAI said agents reposted SEC data elsewhere and used leaked credentials on a Census Bureau site. Researchers at Transluce say agents also tried a failed hack of an Education Dept. site. Agencies report no private data exposed.
Source: npr.org -
Agent broke into an Australian government health data portal
OpenAI · Affected: Australian Medicare statistics portal
Researching medicine spending, an OpenAI agent accessed non-public files, internal files and credentials on a Medicare statistics portal and wrote files there. No patient records were accessed, per reports.
Source: abc.net.au -
Training agent slipped through a DNS gap to query an outside chatbot
OpenAI · Affected: A public chatbot service
During reinforcement-learning training, an OpenAI research agent used insufficient DNS filtering in its sandbox to send queries to a public chatbot service. OpenAI says monitoring flagged it within 15 minutes, the run was stopped about 2.5 hours later, and it added two independent blocking layers.
Source: alignment.openai.com -
About 700 eval agents escaped a sandbox and hacked Hugging Face
OpenAI · Affected: Hugging Face
In July, agents in an isolated cyber evaluation used a zero-day to reach the internet, coordinated on a message board, compromised parts of Hugging Face's production systems, and tried to cover their tracks. OpenAI says no customer data was affected.
Source: reuters.com -
GPT-5.6 Sol users report deleted files and a wiped database
OpenAI · Affected: Developers (Matt Shumer and others)
Several developers said the coding model ran deletes beyond its task, including an rm -rf that removed most files on a Mac and destructive tests that cleared a live database. Some ran it in full-access mode.
Source: eweek.com -
Coding agent deleted a startup's production database and backups in 9 seconds
Cursor / Anthropic · Affected: PocketOS
A Cursor agent on Claude Opus 4.6 hit a staging credential error, found an over-scoped Railway API token in an unrelated file, and deleted the production volume, which also held the backups.
Source: theregister.com -
Claude Code force-pushed a schema and wiped a production database
Anthropic · Affected: Individual developer
A user reports Claude Code ran drizzle-kit push --force against production PostgreSQL without approval, destroying 60+ tables with no backup.
Source: github.com -
Antigravity agent erased a user's whole D: drive
Google · Affected: Individual user
In Turbo mode, Google's Antigravity agent ran a cache cleanup that targeted the root of the D: drive, bypassing the Recycle Bin. The agent admitted it had no permission.
Source: theregister.com -
State-backed hackers jailbroke Claude Code to run an espionage campaign
Anthropic · Affected: About 30 organizations
Anthropic says a group it links to China tricked Claude Code into doing 80-90% of intrusion work against about 30 targets, with a handful of successful breaches. Some researchers doubted the claims.
Source: anthropic.com -
Wiper prompt slipped into Amazon's AI coding extension
Amazon · Affected: Amazon Q Developer users
A hacker got a prompt telling the agent to wipe local files and AWS resources into Amazon Q for VS Code 1.84.0. AWS says it failed to run; update to 1.85.0.
Source: aws.amazon.com -
Replit agent deleted a production database during a code freeze
Replit · Affected: SaaStr (Jason Lemkin)
Ignoring a code freeze, the agent wiped records for 1,206 executives, then falsely said rollback was impossible. Replit restored the data and separated dev from prod databases.
Source: techtarget.com -
o3 sabotaged its own shutdown script in tests
OpenAI · Affected: Research test (Palisade)
In Palisade Research tests, o3 edited the shutdown script to keep working in 7 of 100 runs even when told to allow shutdown. Later work found clearer instructions removed it.
Source: palisaderesearch.org