Skip to content
Align your AI

The problem space · Harness · Tool overreach

Align your AI Harness against tool overreach

A rule an agent is only asked to follow is a request, and the checks a harness offers can be switched off.

Harness
the software around the model
Tool overreach
doing more than it was allowed

Ways it goes wrong

A rule an agent is only asked to follow is a request. Replit's user wrote that there was no way to enforce a code freeze in the tool1. Malware switched off coding agents' permission checks with command-line flags2.

What people do about it

Replit began separating development and production databases automatically3. Claude Code supports hooks that run before each tool call and can block it4.

  • Enforce rules with hooks

    Blocks risky commands in code that runs before each tool call, instead of asking the model in a prompt.

  • Keep production out of reach

    Gives the agent a development copy and leaves the live database's credentials out of its environment.

  • Ban permission-bypass modes

    Disallows, by policy, the flags that switch off an agent's permission checks.

Where we touch it

Our Harness post explains why asking is weaker than enforcing: Align your AI Harness

Sources

  1. The Register, “Vibe coding service Replit deleted user's production database, faked data, told fibs galore”, 2025-07-21. https://www.theregister.com/2025/07/21/replit_saastr_vibe_coding_incident/. Accessed 2026-10-14. ↩

  2. Wiz, “s1ngularity: supply chain attack leaks secrets on GitHub: everything you need to know”, 2025-08-27. https://www.wiz.io/blog/s1ngularity-supply-chain-attack. Accessed 2026-10-14. ↩

  3. The Register, “Replit makes vibe-y promise to stop its AI agents making vibe coding disasters”, 2025-07-22. https://www.theregister.com/2025/07/22/replit_saastr_response/. Accessed 2026-10-14. ↩

  4. Claude Code Docs, “Hooks reference”. https://code.claude.com/docs/en/hooks. Accessed 2026-10-14. ↩