A senior product engineer asked Claude Code to identify easy security improvements in a service they control. The tool found a minor vulnerability that could be fixed with a one-line change, but safeguards blocked further help. Seeking an exemption would mean waiting several business days, surrendering data-retention protections, submitting government identification, and still facing possible rejection. The engineer says attempts to investigate real threats - including whether public exploit code could be adapted to attack their technology stack - were likewise refused, even when using older models.
The writer argues that these restrictions prevent legitimate defenders from using AI to secure their own systems, while criminals can turn to unrestricted models and face no comparable approval process. In this case, EU-hosted open-weight models, including GLM 5.3, were able to confirm the vulnerability, help fix it, and verify the repair. The experience leads to a broader criticism: AI companies apply limits that obstruct defensive work without stopping malicious use, potentially making security worse. The writer compares AI tools to knives, which can be dangerous but are not subject to identity checks, and speculates that restrictions may benefit major labs by encouraging regulation that disadvantages competitors.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.