November 20, 2026AgentsCoding

A Developer Hits Claude's Cyber Gate and Walks to GLM

The most upvoted complaint about Claude this week is not about quality. It is about being told no. A short, very angry blog post from a senior product engineer (https://ounapuu.ee/posts/2026/11/19/unsafeguards/, published November 19, 04:00 UTC) hit 389 points on Hacker News within hours. We are not repeating the title.

The story is simple. The author asked Claude Code to look for low-hanging security fixes in a product they control. It found an easy one-line oversight, then the session got stopped by Anthropic's cyber safeguards. The route around it is the Cyber Verification Program, and the author lists what that costs: a wait of what they describe as multiple business days, identity verification with an ID or passport, giving up data-retention exemptions, and no guarantee of approval. Their workaround was an EU-hosted open-weight model, GLM 5.3, which confirmed the issue and helped fix and verify it. The punchline: attackers already use models with the safeguards stripped out, so the gate only slows down defenders.

Anthropic's own page (https://support.claude.com/articles/14604842-cyber-verification-program) frames it the other way, and it matters what it actually says. It states that you do not need CVP for everyday secure development, and that generally available models can review code, patch known issues and triage alerts. CVP is meant for things like validating exploits and authorized penetration tests, with three tiers up to critical-infrastructure red teaming reviewed with the US government. Data retention is required for everyone in the program so Anthropic can watch for misuse. By that description, a one-line defensive fix should not have been blocked at all. So either the classifier misfired, or the line between "check whether a public proof of concept works against my stack" and "build an exploit" is blurrier in practice than on the policy page. The author's post suggests the second.

Both sides have a point, and that is the uncomfortable part. Labs gate dual-use security work because the same capability that validates a fix can weaponize a bug, and this month's string of outages has been blamed, by some, on rushed patching. But every false positive sends a defender to an ungated model, and that model does not log anything for anyone. A safeguard that pushes the good-faith user off the platform is measuring its own block rate, not its effect on safety. The number worth publishing is how often the gate stops legitimate defensive work, and nobody publishes it.
← Previous
Three Papers, One Bill: Where Agent Tokens Actually Go
Next β†’
Meta Muse Told an Employee Someone Else's Bank Balance
← Back to all articles

Comments

Loading...
>_