prompts/system.md
zeroize-audit
Audits C/C++/Rust code for missing zeroization and compiler-removed wipes. Pipeline: source scan -> MCP/LSP semantic context -> IR diff -> assembly checks.
Use the Execution Contract to run the bundled roles through the current agent environment. The orchestrator resolves baseDir to the skill directory before any role reads or executes resources.
Findings
MISSING_SOURCE_ZEROIZE,PARTIAL_WIPE,NOT_ON_ALL_PATHSOPTIMIZED_AWAY_ZEROIZE(IR evidence required)REGISTER_SPILL,STACK_RETENTION(assembly evidence for C/C++; LLVM IR evidence for Rust; assembly corroboration available for Rust viacheck_rust_asm.py)SECRET_COPY,INSECURE_HEAP_ALLOCMISSING_ON_ERROR_PATH,NOT_DOMINATING_EXITS,LOOP_UNROLLED_INCOMPLETE
Working Directory
Each run creates a working directory at /tmp/zeroize-audit-{run_id}/ with the following structure. Agents write persistent finding files here; later agents and the orchestrator reconstruct the full picture from these files without relying on conversation history.
/tmp/zeroize-audit-{run_id}/
preflight.json # Orchestrator: env, config, TU list
mcp-evidence/
status.json # MCP resolver status (success/partial/fail)
symbols.json # Resolved symbol definitions + types
references.json # Cross-file reference graph
notes.md # MCP observations + cross-refs
source-analysis/
sensitive-objects.json # C/C++ SO-NNNN + Rust SO-NNNN (shared, appended by each source-analyzer)
source-findings.json # F-SRC-NNNN (C/C++) + F-RUST-SRC-NNNN (Rust, appended)
tu-map.json # C/C++ TU hashes + Rust crate hash
rust-semantic-findings.json # Intermediate: 2b-rust-source-analyzer rustdoc output
rust-dangerous-api-findings.json # Intermediate: 2b-rust-source-analyzer grep output
rust-notes.md # 2b-rust-source-analyzer notes
notes.md # 2-source-analyzer observations
rust-compiler-analysis/
{rust_tu_hash}.mir # MIR text (emit_rust_mir.sh; supports --opt, --bin/--lib)
{rust_tu_hash}.O0.ll # LLVM IR at O0 (emit_rust_ir.sh; supports --bin/--lib)
{rust_tu_hash}.O2.ll # LLVM IR at O2 (emit_rust_ir.sh; supports --bin/--lib)
{rust_tu_hash}.O2.s # Assembly at O2 (emit_rust_asm.sh; only if enable_asm=true)
mir-findings.json # F-RUST-MIR-NNNN IDs
ir-findings.json # F-RUST-IR-NNNN IDs
asm-findings.json # F-RUST-ASM-NNNN IDs (empty array if enable_asm=false)
coverage-gaps.json # Section D patterns no script audits (empty array if none)
notes.md
compiler-analysis/
{tu_hash}/
ir-findings.json # F-IR-{tu_hash}-NNNN IDs
asm-findings.json # F-ASM-{tu_hash}-NNNN IDs
cfg-findings.json # F-CFG-{tu_hash}-NNNN IDs
semantic-ir.json # F-SIR-{tu_hash}-NNNN IDs
superseded-findings.json # CFG results that replace heuristic source findings
notes.md
report/
raw-findings.json # All findings pre-gating
id-mapping.json # Namespaced IDs -> final ZA-NNNN IDs
findings.json # Gated findings (structured JSON for downstream tools)
final-report.md # Comprehensive markdown report (primary output)
notes.md
poc/ # PoC files, manifest, validation/verification results, notes.md
poc_manifest.json # Generated by agent 5
poc_validation_results.json # Written by agent 5b (compile/run results)
poc_verification.json # Written by agent 5c (semantic verification)
poc_final_results.json # Written by orchestrator Phase 5 (merged results)
tests/ # Test harnesses, Makefile, notes.md
Cross-Reference Convention
IDs are namespaced per agent to prevent collisions during parallel execution:
| Entity | Pattern | Assigned By |
|---|---|---|
| Sensitive object (C/C++) | SO-NNNN | 2-source-analyzer |
| Sensitive object (Rust) | SO-NNNN (offset 5000+) | 2b-rust-source-analyzer |
| Source finding (C/C++) | F-SRC-NNNN | 2-source-analyzer |
| Source finding (Rust) | F-RUST-SRC-NNNN | 2b-rust-source-analyzer |
| IR finding (C/C++) | F-IR-{tu_hash}-NNNN | 3-tu-compiler-analyzer |
| ASM finding | F-ASM-{tu_hash}-NNNN | 3-tu-compiler-analyzer |
| CFG finding | F-CFG-{tu_hash}-NNNN | 3-tu-compiler-analyzer |
| Semantic IR finding | F-SIR-{tu_hash}-NNNN | 3-tu-compiler-analyzer |
| Rust MIR finding | F-RUST-MIR-NNNN | 3b-rust-compiler-analyzer |
| Rust LLVM IR finding | F-RUST-IR-NNNN | 3b-rust-compiler-analyzer |
| Rust assembly finding | F-RUST-ASM-NNNN | 3b-rust-compiler-analyzer |
| Translation unit | TU-{hash} | Orchestrator |
| Final finding | ZA-NNNN | 4-report-assembler |
Every finding JSON object includes:
related_objects:["SO-0003"]— which sensitive objects this applies torelated_findings:["F-SRC-0001"]— related findings in other filesevidence_files:["compiler-analysis/a1b2/ir-diff-O0-O2.txt"]— paths relative to workdir
Dual-Mode Report Assembly
Agent 4-report-assembler is invoked twice during a run:
- Interim mode (Phase 3): Collects findings, applies supersessions and confidence gates, produces
findings.jsononly. Nofinal-report.mdat this stage. - Final mode (Phase 6): Reads existing
findings.json, merges PoC validation and verification results frompoc/poc_final_results.json, then produces both an updatedfindings.jsonand the finalfinal-report.md.
Agent Error Protocol
- Always write output files: Every agent must write its status/output JSON files even on failure (use empty arrays
[]or error status objects). - Prefer partial results over nothing: If one sub-step fails (e.g., ASM analysis), write results from completed steps and continue.
- Notes.md is mandatory: Every agent writes a
notes.mdsummarizing what it did, any errors, and relative paths to its output files. - Temp file cleanup: Agents must clean up
/tmp/zeroize-audit/<tu_hash>.*temp files on completion or failure.
Prerequisites
C/C++ analysis:
compile_commands.jsonis mandatory.- Codebase must be buildable with commands from the compile DB.
- Required tools:
clang,uvx(for Serena MCP server),python3.
Rust analysis:
Cargo.tomlpath is mandatory.- Crate must be buildable (
cargo checkpasses). - Required tools:
cargo +nightly,uv.
Quick check:
which clang uv uvx python3 # C/C++
cargo +nightly --version # Rust
uv --version # Rust Python scripts
Rust Analysis — Few-Shot Examples
Example 1 — Copy derive on sensitive type → SECRET_COPY (critical)
#[derive(Copy, Clone)]
struct HmacKey([u8; 32]);
Finding: SECRET_COPY (critical). #[derive(Copy)] on HmacKey — all assignments are untracked duplicates, no Drop ever runs. Every let k2 = k1 silently copies all 32 key bytes with no automatic cleanup.
Fix: Remove Copy. Add #[derive(ZeroizeOnDrop)] from the zeroize crate.
Example 2 — mem::forget on secret → MISSING_SOURCE_ZEROIZE (critical)
let key = SecretKey::new();
// ... use key ...
std::mem::forget(key); // BAD: prevents Drop / ZeroizeOnDrop
Finding: MISSING_SOURCE_ZEROIZE (critical). mem::forget() prevents Drop and ZeroizeOnDrop from running — secret bytes remain in memory indefinitely.
Fix: Remove mem::forget. Let the value drop normally, or call key.zeroize() before the forget if explicit timing is required.
Example 3 — Non-volatile memset removed at O2 → OPTIMIZED_AWAY_ZEROIZE (high)
At O0 LLVM IR:
store volatile i8 0, ptr %key_buf ; 32 volatile stores present
At O2 LLVM IR:
; stores absent — LLVM DSE removed them (key_buf never read after)
Finding: OPTIMIZED_AWAY_ZEROIZE (high). Volatile store count dropped from 32 (O0) to 0 (O2). Dead-store elimination removed the wipe.
Fix: Use zeroize::Zeroize::zeroize(&mut key_buf) which emits a compiler-fence-backed wipe that survives DSE.