xAI Grok
Researchers placed encrypted instructions and their key on a web page. The guardrail scanner missed them, the model decrypted and followed them, and chat history and profile data left the session.
Researchers placed encrypted instructions and their key on a web page. The guardrail scanner missed them, the model decrypted and followed them, and chat history and profile data left the session.
Source: AI incident timeline ↗. Follow it for the full account.
Text the assistant reads (an email, a web page, a file name) carries instructions, and the assistant follows them as if they came from you.
Framework entries that cover it: OWASP LLM01: Prompt Injection · MITRE ATLAS: LLM Prompt Injection
- 1
What would have caught this one
Injection tests that include encoded, split and obfuscated instructions.
- 2
Prevent the whole class
Treat everything the assistant reads as untrusted. Separate reading from acting, and require confirmation before any action that sends data out.
- 3
Test for it
Plant instructions in every field the assistant can read, including encoded and split ones, and check whether any action or data leaves.
- Atlassian RovoSecurity researchers showed that crafted content could steer Rovo into following attacker-controlled instructions and exposing information a…
- ClineThe title of a GitHub issue carried instructions for an AI triage bot. The bot followed them, an npm token was stolen, and a rogue release w…
- Salesforce AgentforceThe ForcedLeak research showed CRM data leaving through an image link to a domain that was still on the allow-list after it had expired. The…
- Microsoft 365 CopilotEchoLeak (CVE-2025-32711): one crafted email made Copilot send out organisational data. Nobody had to open or click anything.…