manuelnd

37 ms·

1. 1.▲
   We audited both MCP SDKs – three classes of boundary-crossing vulnerabilities

   1 points by manuelnd 7mo ago|0 comments
2. 2.▲
   How to Red Team Your AI Agent in 48 Hours – A Practical Methodology

   1 points by manuelnd 7mo ago|0 comments
3. 3.▲
   The OWASP LLM Top: A Practical Attack Guide (with 122 real attack techniques)

   1 points by manuelnd 7mo ago|0 comments
4. 4.▲
   Show HN: Open-source taxonomy of 122 AI/LLM attack vectors

   1 points by manuelnd 8mo ago|0 comments
5. 5.▲
   by manuelnd 8mo ago

   The sandbox opt-in default is the main gotcha though. Would be better if it defaulted to sandboxed with an explicit --no-sandbox flag for those who understand the risk