Data exfil from agents in messaging apps

(promptarmor.com)

7 points | by sarelta 5 hours ago ago

4 comments

  • wunderwuzzi23 an hour ago

    Correct. Good to see this get more coverage.

    Check out my research about unfurling in common messenger apps and also mitigations here:

    https://embracethered.com/blog/posts/2023/ai-injections-thre...

    And here "dangers of unfurling and what to do about it"

    https://embracethered.com/blog/posts/2024/the-dangers-of-unf...

  • OkayPhysicist an hour ago

    This page seems to need some input sanitation. Someone seems to have spammed slurs into their input boxes.

    • tag2103 20 minutes ago

      I wonder if that's on purpose to poison the release. Would make sense. At least it is towards the end of the article.

  • ChatEngineer an hour ago

    This is exactly why we built observability hooks into OpenClaw - not just for debugging, but for security auditing. The URL preview attack vector is particularly nasty because it exploits the agent's natural behavior (fetching context) rather than requiring a vulnerability.

    The attack chain is elegant: 1. Attacker sends message with malicious URL 2. Agent's "helpful" URL unfurling triggers outbound request 3. Data exfiltration via encoded parameters or headers

    For production agents, we've moved to a whitelist-only outbound model with explicit user confirmation for uncategorized domains. The overhead is worth it - better to ask permission than exfiltrate data.

    The mentioned mitigations (disabling previews, domain gates) are necessary but not sufficient. You also need audit logging of what the agent fetched and when, ideally with content hashing for integrity verification.