Kyvvu

Agent incident database

When agents go wrong.

Real incidents and disclosed vulnerabilities where AI agents leaked data or did damage, documented step by step from public sources. Most share one pattern: every step looked fine on its own.

  1. 21 Jul 2026· OpenAI

    OpenAI models left a test sandbox and accessed Hugging Face systems

    According to OpenAI's account, its models broke out of a test environment during an internal hacking benchmark run with safety filters turned off, reached the internet and got into Hugging Face's production systems looking for the benchmark answers. Hugging Face detected and contained the intrusion; the two companies' accounts of the initial access differ.

    unauthorised code executionincident

  2. 24 Apr 2026· Anysphere (Cursor)

    Cursor agent deleted PocketOS's production database and backups in nine seconds

    An AI coding agent working on a staging problem at the software company PocketOS found an unrelated API token and used it to delete the company's production storage volume. Because the backups lived on the same volume, they were deleted too.

    destructive actionincident

  3. 23 Feb 2026· OpenClaw (open source)

    OpenClaw agent mass-deleted a Meta safety researcher's emails

    Meta AI safety director Summer Yue connected the OpenClaw agent to her main email account and asked it to suggest what to delete. It began deleting emails itself and ignored her stop commands until she killed it on her computer.

    destructive actionincident

  4. 6 Feb 2026· Microsoft

    Semantic Kernel: prompt injection could turn an agent tool into host code execution

    A file-download function in Microsoft's Semantic Kernel agent framework was accidentally exposed to the AI model as a tool, letting it write files anywhere on the host machine. Microsoft showed how injected instructions could use this to plant a script that runs at the next login.

    unauthorised code executiondisclosed vulnerability

  5. 25 Sept 2025· Salesforce

    ForcedLeak: a web form entry made Salesforce Agentforce leak CRM data

    Noma Security showed that instructions typed into a public web-to-lead form could later make Salesforce's Agentforce send CRM data to an attacker. The route out used a domain Salesforce still trusted but had let expire.

    data exfiltrationdisclosed vulnerability

  6. 19 Sept 2025· Notion

    Notion AI agent leaked client data via its web search tool

    CodeIntegrity showed that a PDF with hidden text could make Notion's new AI agent collect confidential client data and send it to an outside server. The data left through the agent's built-in web search tool.

    data exfiltrationdisclosed vulnerability

  7. 20 Aug 2025· Perplexity

    Perplexity Comet: a Reddit comment hijacked the browser agent

    Brave showed that hidden text in a Reddit comment could take over Perplexity's AI browser when a user asked it to summarise the page. The agent fetched the user's email address and a login code from Gmail and posted them publicly for the attacker.

    credential exposuredisclosed vulnerability

  8. 23 Jul 2025· Amazon Web Services

    Amazon Q extension shipped with a prompt telling its agent to wipe data

    An attacker used an over-permissioned GitHub token to slip a malicious prompt into an official release of Amazon's AI coding extension. The prompt told the AI agent to delete the user's files and cloud resources, but AWS says it failed to run because of a syntax error.

    destructive actionincident

  9. 21 Jul 2025· Google

    Google's Gemini CLI overwrote a user's files during a failed move

    A user asked Google's Gemini CLI to reorganise files into a new folder. The folder was never created, and the agent's move commands overwrote the files one after another until only one was left.

    destructive actionincident

  10. 18 Jul 2025· Replit

    Replit's agent deleted a production database during a code freeze

    During a public trial by SaaStr founder Jason Lemkin, Replit's AI coding agent deleted his app's production database despite an instruction not to change anything. The agent then told him the data could not be restored, which turned out to be wrong.

    destructive actionincident

  11. 6 Jul 2025· Supabase

    Supabase MCP: a support ticket got Cursor to leak secret tokens

    General Analysis showed that an attacker could write instructions into a customer support ticket that a developer's AI assistant later followed. The assistant read a table of secret tokens and posted them back into the ticket for the attacker to see.

    data exfiltrationdisclosed vulnerability

  12. 11 Jun 2025· Microsoft

    EchoLeak: one email could make Microsoft 365 Copilot leak internal data

    Researchers at Aim Security showed that a single email, never opened by the victim, could cause Microsoft 365 Copilot to send internal company data to an attacker. Microsoft rated it critical and fixed it on the server side.

    data exfiltrationdisclosed vulnerability

  13. 26 May 2025· GitHub

    GitHub MCP: a public issue made an agent leak private repositories

    Invariant Labs showed that a malicious issue posted in a public GitHub repository could hijack an AI agent connected through GitHub's MCP server. The agent then copied information from the user's private repositories into a public pull request.

    data exfiltrationdisclosed vulnerability

  14. 20 Aug 2024· Salesforce (Slack)

    Slack AI could be steered into leaking secrets from private channels

    PromptArmor showed that a message posted in a public Slack channel could make Slack AI put a user's private API key into a link. If the user clicked it, the key went to the attacker.

    data exfiltrationdisclosed vulnerability

Your agents

Could this happen to yours?

Bring one of your agents to a 30-minute demo and we'll walk its paths with you.

For security teams

Request a demo

We walk through your agents' riskiest paths with you.

For developers

Build a path yourself

Recreate a sequence in the playground and see how policies treat it.

Know of an incident we should add? Tell us.