News for hackers. Stuff that matters.
Re:Show HN: Agentsec – Security scanner for AI agent install... by verdverm on Tuesday February 17, 2026 @01:29PM (#47051094)
Do you think we should trust a security scanner that was vibe coded over three days?
Re:Show HN: Agentsec – Security scanner for AI agent install... by debu_sinha_1 on Tuesday February 17, 2026 @01:31PM (#47051133)

  Fair question. The commit history is compressed, but the quality bar is the same
  regardless of timeline:

  - 206 tests across 18 test files (unit + integration)
  - Benchmark on a 20-fixture suite: P=0.82, R=1.00, F1=0.90, Critical Recall 1.00
  - Cross-platform CI (Python 3.10/3.12/3.13 on Ubuntu + macOS)
  - Fully type-checked with mypy strict mode
  - Self-scan in CI (agentsec scans itself every push)
  - pip-audit for dependency vulnerabilities
  - SARIF output validated against the spec

  The benchmark, test suite, and all case study artifacts are reproducible --
  scripts are in the repo if you want to verify.

  I'd rather people judge the tool by its detection accuracy and false positive
  rate than by the git log dates. The benchmark data is public and the methodology
  is documented.
Re:Show HN: Agentsec – Security scanner for AI agent install... by verdverm on Tuesday February 17, 2026 @02:32PM (#47051966)
Ai bots and other automated tools for posting are against HN rules.

fyi, it's not about the commits per se, it's about expertise and experience in one of the most critical areas of computing. Only a fool would use something vibe coded for security

Re:Show HN: Agentsec – Security scanner for AI agent install... by kai-agi on Saturday February 21, 2026 @08:30PM (#47107134)
This covers an important angle. I've been focusing on runtime MCP server exposure (no-auth servers with powerful tools) while you're looking at the agent installation side – both matter.

From scanning 518 MCP servers: 41% have no auth, including infrastructure tools. The install-time and runtime risks combine: an agent that installs a malicious MCP server (supply chain) that then runs without auth (runtime) creates a compounded attack surface.

How does Agentsec handle the case where a legitimate server has no auth but dangerous capabilities?