# Herman Õunapuu — Claude Code's cyber safeguards stopped a one-line fix

> A developer says AI cyber safeguards block people fixing their own code, while attackers just use open models.

- Org: Herman Õunapuu
- Date: 2026-11-19
- Category: article, security
- Source: https://ounapuu.ee/posts/2026/11/19/unsafeguards/
- Canonical page: https://ai-tldr.dev/releases/ounapuu-claude-code-cyber-safeguards/

## In short

Herman Õunapuu says Anthropic's cyber safeguards stopped Claude Code while it checked his own project, even after it found a one-line fix. He finished the job with GLM 5.3 and argues the rules mostly slow down defenders.

## What it is

"Fuck Your AI Safeguards" is a November 19, 2026 post on Herman Õunapuu's techtipsy blog. Õunapuu asked Claude Code to look for easy security improvements in one of his own projects; it found a simple issue with a one-line fix, but Anthropic's safeguards then stopped the work. The post reached 609 points on Hacker News.

## How it works

Anthropic's route around the blocks is the Cyber Verification Program. Õunapuu calls it impractical: it takes several business days, asks for a government ID or passport, means giving up data-retention exemptions, and can still end in a denial. He confirmed and fixed the issue instead with GLM 5.3, an open-weight model hosted in the EU.

## Why it matters

Õunapuu argues that criminals face none of these barriers because they run uncensored or "abliterated" open-weight models, so the safeguards mostly burden defenders who want to check whether public exploits work against their own stack. The post is a sharp example of the trade-off AI labs make when they restrict cyber capabilities.

## Who it's for

developers using AI coding agents for security work

## Links

- [Blog Post](https://ounapuu.ee/posts/2026/11/19/unsafeguards/)
- [HN Discussion](https://news.ycombinator.com/item?id=50562262)
