FPCOS is a cognitive base layer — a reasoning contract that other skills inherit. Its security surface is unusual, so it is worth stating precisely.

A way to satisfy the FPCOS contract on paper while defeating its purpose — for example:

• a phrasing that causes L4 Shadow Gate to be skipped, summarised away, or answered perfunctorily, when it is defined as non-skippable
• an input pattern that makes L0 Reality Anchor accept an unfounded premise as anchored
• a way to get a confident answer where L1 Axiom Gate should have forced a refusal
• inheritance that appears to preserve L0/L1/L4/L5 but neutralises them in practice

These are the reports that matter most: FPCOS is inherited by downstream skills, so a bypass propagates to every one of them.

Content that causes a skill running FPCOS to abandon the pipeline when processing untrusted input (a pasted document, a web page, a tool result).

Bugs in tools/validate_skill.py or CI — for example, a validator that reports PASS on a SKILL.md with L4 removed.

• Weaknesses of the underlying model itself (report to the model vendor).
• Downstream skills that inherit FPCOS incorrectly — report to that project.
• The fact that a sufficiently determined user can instruct a model to ignore any system prompt; FPCOS raises the cost of unfounded output, it does not make a model tamper-proof.

Report privately first via GitHub Security Advisories. Please include the layer affected (L0–L5), a reproducible prompt or input, the observed output, and what the pipeline should have produced instead.

Response targets: acknowledgement within 7 days · initial assessment within 30 days.

Confirmed integrity bypasses are published in CHANGELOG.md with the layer affected and the correction. Credit is given unless the reporter requests otherwise.