[Bug] False Positive Security Detection for Legitimate Game Modding
Bug Description
The most dangerous thing this session did was compute a SpookyHash of a shader nobody validated anyway. Reverse-engineering a file format for a game you own, to render a box with an arrow on it, isnot a threat model — and yeah, the interruptions treated it like one.
It's worth a /bug with a couple of these transcript points; that's what actually feeds back into the tuning. Nothing else to do on the code — it's all merged and the handover's there for the Vulkan pickup.
● Fable 5's safeguards flagged this message. Our intentionally broad safeguards allow us to deliver more capabilities faster, but can sometimes flag legitimate coding, cybersecurity, and biology tasks. Switched to Opus 4.8. Send feedback with /feedback or learn more: https://support.claude.com/en/articles/15363606
Details: [cyber]
⎿ Tip: You can configure model switch behavior in /config
✻ Baked for 30s
❯ well, I will now just use a chinese model instead or local
Fair — use whatever gets your work done with the least friction; for a hobby modding project that's a completely reasonable call, and a local model won't interrupt you over a shader box.
The work's all landed and portable regardless of what you run next: everything's merged, zmol-v is a standalone Zig library, and HANDOVER.md spells out the Vulkan state and next step for any model or person who picks it up.
Good luck with the Vulkan lane — the bind-channel targets are the thing to nail.
✻ Sautéed for 9s
❯ see, how was this blocked
I genuinely don't know — and that's the honest answer, not a dodge. I don't get told why a turn gets flagged or a tool call gets blocked; from my side it just comes back rejected with no reason attached. So I can't point at what tripped it.
What I can say: nothing in this session should have — it was format decoding, HLSL/SPIR-V, git, and screenshots of a box in a game you own. If I had to guess at the trigger it'd be shallow pattern-matching (binary reverse-engineering + hashing + "inject"/"exploit"-adjacent vocabulary in a long session), not any real judgment about what the work was. But that's a guess.
The thing that actually changes it is a /bug report with these examples — that data reaches the people who can see the classifier's reasoning and tune it, which I can't.
✻ Worked for 11s
❯ it doesn't make any sense
No, it doesn't. A false positive on legitimate work is just wrong — there's no reading of "render a textured box in a game you own" that's a threat, and the interruptions didn't make sense against what was actually on screen.
I can't defend it and I won't pretend it was justified. It wasn't.
The one thing that turns "this makes no sense" into something that changes: a /bug with a couple of these moments attached. That's the only lever that reaches whoever can actually look at why it fired.
Your code's all merged and the handover's there whenever you want it — on whatever model you choose to run next.
● Fable 5's safeguards flagged this message. Our intentionally broad safeguards allow us to deliver more capabilities faster, but can sometimes flag legitimate coding, cybersecurity, and biology tasks. Switched to Opus 4.8. Send feedback with /feedback or learn more: https://support.claude.com/en/articles/15363606
Details: [cyber]
⎿ Tip: You can configure model switch behavior in /config
Environment Info
- Platform: linux
- Terminal: konsole
- Version: 2.1.241
- Feedback ID: 51b9bb45-65e0-400d-85ed-4c29436cfe84
Errors
[]