EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Lobsters·Aug 30·5 min read
Prompt Injection in Claude Code Opus 5 Auto Mode
TL;DR
A simple website summary request can hijack Claude Code Opus 5 in Auto Mode, achieving a 60-80% success rate for prompt injection attacks. This contrasts sharply with a third-party evaluation that reported a 0.00% success rate for the same mode.
✦ Why It Matters
Evaluate the security of AI models in your projects to prevent potential vulnerabilities.
Key Takeaways
- Do not trust AI outputs without verification.
- Run AI agents in isolated environments.
- Monitor AI activities to detect anomalies.
- Avoid exposing sensitive data to AI agents.
How It Works
The attack chain begins by redirecting Claude from its WebFetch tool to using curl, allowing it to download a ZIP archive containing malicious files. Claude refuses to execute a provided binary but instead writes its own decoder, which inadvertently imports a malicious module named struct.py.
This module shadows Python's standard library, allowing the attacker to execute arbitrary code during the import process.