Fuck Your AI Safeguards
- ID
- 48508
- Status
- summarized
- Published
- 20 Nov 2026, 12:30 AM
- Fetched
- 20 Nov 2026, 6:32 AM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://ounapuu.ee/posts/2026/11/19/unsafeguards/
- Source URL
- https://hnrss.org/best
Summary
- Score
- 7.5
- Created
- 20 Nov 2026, 6:33 AM
- Tags
- Audience
- developersvibe_codersai_ml_learnerssaas_founders
What happened
A senior product engineer describes asking Claude Code to look for low-hanging security fixes on their own product; it found a one-line oversight, then refused to help apply it because of "AI safeguards." The exemption route described in the post requires waiting multiple business days, forfeiting data-retention exemptions, and submitting an ID or passport, with no guarantee of approval. They ultimately used an EU-hosted open-weights model, GLM 5.3, to confirm the issue and verify the fix. The thread drew 382 points and 153 comments on Hacker News.
Why it matters
If your workflow depends on a hosted frontier model for defensive security work, this is a concrete failure case: a one-line fix got blocked, and the documented escape hatch costs business days plus an ID document and your data-retention exemption. Decide now which open-weights model you can run or rent regionally as a fallback, and test it on a real defensive task before you need it, rather than discovering the wall mid-incident. The claim that abliterated open-weights models are available to attackers regardless is the part worth weighing against the compliance cost.
Discussion angle
Run the test live: give your coding agent a genuine defensive task on your own repo and see where it refuses — then ask who in the room has a tested open-weights fallback, and what it costs them in latency, data residency, and compliance paperwork.