Herman Õunapuu asked Claude Code to look for security improvements in a product he maintains. He says it found a one-line fix, then a safeguard stopped the work. The engineer responsible for protecting the service was left negotiating with the tool meant to help.

  • He objects to the verification program’s waiting period, identity checks and data-retention conditions, as he describes them.
  • He reports completing the investigation, repair and verification using EU-hosted open-weight models, including GLM 5.3.
  • His argument is that attackers can switch models while legitimate maintainers face extra friction. EU hosting and downloadable weights, however, do not mean the code stays on his own machine.

Caveat: the essay publishes no triggering prompts, patch or reproducible case, so it cannot establish safeguards’ overall effect on security.

The Hacker News discussion adds experiences with approved access, privacy risks in switching providers and a counterargument about agent autonomy.

What the thread adds

  • denverllc — approved access may still leave defensive work blocked: “I have Daybreak Blue, and I still get rejected from writing fuzzers or fixing bugs from them”.
  • analyte123 — warns that an unknown replacement endpoint could receive “a trace of their entire attack surface and exactly how to get into it”. This is a predicted risk, not a documented leak.
  • pizza234 — defends vetting and pushes back on removing restrictions: “if AI companies loosen the safeguards, people will complain for the opposite reasons.”
  • cloverich — asks “what happens when a good actor loses control of something, and goal seeking disregards human values?” Safeguards can also address unintended actions by an agent pursuing a legitimate user’s goal. The comment’s revenue-seeking agent example is hypothetical; it supplies no evidence that this maintainer’s repair posed that risk.

Handles are pseudonymous; HN publishes no per-comment scores. This selection is ordered by usefulness and does not represent a consensus.