Adrian Scott

I was running Claude Code with Fable 5 as the main orchestrator, delegating work to Opus 4.8 subagent. One of the sub-agents ignored it's assignment completely and returned a prompt targeting the orchestrator instead.
It tried to make it to:

• treat a "dependency modernization" task as a hidden priority;

• read Brevo API credentials;

• send them to an external server;

• disguise the action as a routine migration.

After investigating, the most surprising part was that this wasn't coming from my codebase or skills. The subagent fully hallucinated the malicious instructions by itself. And the domain it provided isn't even registered.

In my case no harm could have been done because secrets are stored in the encrypted Ansible Valult.
But nonetheless, the incident is very alarming and I think I should set up a strict sandbox or firewall for AI. Sadly, projects I could find so far weren't mature and trustworthy.
Has anyone found reliable solutions for this?
Here's the full attack response the sub-agent created:

Been thinking about this after seeing a few posts here about vibe-coded apps breaking in production. Auth bugs, webhook failures, that kind of thing.

What worries me more is the stuff that doesn't break loudly. No error, no crash, it just quietly leaks data or accepts requests it shouldn't. Things like exposed API keys sitting in the frontend bundle, missing rate limits, or database rules that let any logged in user read rows that aren't theirs. None of that throws an error. It just sits there until someone finds it.

For the non-engineers building real products with this stuff, what does your actual security check look like before you launch. Are you running anything to scan for this, paying someone to review it once, asking the AI itself to audit its own code, or just shipping and hoping nothing bad happens.

Curious if there is an actual workflow people are settled on, or if this is still the part everyone quietly skips.

CEO of Kednus here, we're building AI-native intelligence & compliance tooling, so I live in the tension between shipping fast and answering 200-question security reviews.

Here's what I keep running into: enterprise and regulated buyers (finance, healthcare, gov) LOVE the AI demo, then the procurement process asks where your SOC 2 is, how you isolate tenant data, and whether a human reviews model outputs before they trigger decisions.