cupcake security-review.instructions.md

Code-review instructions for Cupcake, a security and governance layer for AI coding agents. It checks an agent’s tool requests against rules and can allow, block, warn, or ask for approval.

In plain words
What is it for?
Reviewing changes to Cupcake’s policy engine, agent integrations, event handling, and security-hardening branch. OPA Rego is the rule language used to define its policies.
Why use it?
They focus review on security fixes and on how Cupcake handles different AI-agent platforms. This helps identify bypasses and other weaknesses in policy enforcement.

Instructions file for GitHub Copilot

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add instructions/eqtylab/cupcake/security-review
Clone the repo
git clone --depth 1 https://github.com/eqtylab/cupcake

Made for: GitHub Copilot.

Per session 6,351 This file is loaded in full into every session.
When invoked 6,351 The same file — it is already loaded in full.
Security scan B 1 finding. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.06351 $0.06351
Opus 5 $0.03175 $0.03175
Sonnet 5 $0.01270 $0.01270
Haiku 4.5 $0.00635 $0.00635

Measured 2d ago against content hash 29ee67a48d3c, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade B, and why

cupcake security-review.instructions.md scanned grade B with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Recursive force deletemediumDestructive command

rm -rf with a variable or a broad path is one typo away from removing the wrong tree.

input.tool_input.command contains "rm -rf"

Downgraded: this mod is about security review, or the phrase is quoted, so it is likely naming the pattern rather than instructing it.

.github/instructions/security-review.instructions.md · 779 lines

How it starts

The opening of the file, as written. The whole thing — 779 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Custom Review Instructions for GitHub Copilot AI Reviewer

Branch Overview

Branch: tob/fix-bypasses Target: main Commits: 45 commits Purpose: Address external security audit findings from Trail of Bits (TOB) and implement comprehensive defense-in-depth security fixes

This branch represents a major security hardening effort implementing fixes for critical vulnerabilities identified in a professional security audit of the Cupcake policy engine.


What is Cupcake?

Cupcake is a policy enforcement engine for AI coding agents (like Claude Code and Cursor). It acts as a security and governance layer that:

  1. Intercepts tool calls from AI agents before they execute
  2. Evaluates them against policies written in OPA Rego (compiled to WebAssembly)
  3. Returns decisions (Allow/Block/Warn/Ask) to enforce rules

Core Architecture:

AI Agent → Hook Event → Cupcake Engine → Policy Evaluation (WASM) → Decision → AI Agent

Key Concepts

1. Harnesses

A harness is the AI coding agent platform (Claude Code, Cursor, etc.). Each harness:

  • Has unique event formats (different JSON structures)
  • Has unique response formats
  • Has different capabilities (context injection, file access)

Critical Design Decision: Cupcake uses a harness-specific architecture with NO normalization layer. Events flow through in their native format, and policies are physically separated by harness:

.cupcake/policies/
├── claude/          # Claude Code policies
└── cursor/          # Cursor policies
2. Hooks

Hooks are integration points where AI agents call external commands during their execution. Examples:

  • PreToolUse - Before a tool executes (e.g., before running a shell command)
  • PostToolUse - After a tool completes
  • UserPromptSubmit - Before processing user input
  • SessionStart - When a session begins

Cupcake integrates via hooks in .claude/settings.json or .cursor/settings.json:

{
  "hooks": {
    "PreToolUse": [{
      "hooks": [{
        "type": "command",
        "command": "cupcake eval --harness claude"
      }]
    }]
  }
}

Read the full file on GitHub · 779 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 2d ago First seen · 779 lines · 6,351 tokens per session scan B 29ee67a48d3c

Subscribe to this mod's changes

cupcake security-review.instructions.md is an instructions file published in the GitHub repository eqtylab/cupcake (288 stars, last pushed 6mo ago), licensed Apache-2.0. It adds 6,351 tokens to every session, about $0.0318 per session on Opus 5. A static security scan graded it B with 1 finding (recursive force delete). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.