trailmark-finding-triage

Files (8)

• SKILL.md
• openai.yaml · agents/
• trail-of-bits-mark.svg · assets/
• input-normalization.md · references/
• output-format.md · references/
• query-recipes.md · references/
• evidence_packet.py · scripts/
• test_evidence_packet.py · scripts/

SKILL.md (first lines)

# Trailmark Finding Triage

Build a concise graph evidence packet for one candidate finding. This skill
answers whether the affected code is reachable, what graph evidence supports
or weakens the claim, and what manual review is still required before calling
the issue exploitable.

## When to Use

- Triage one static-analysis result before spending PoC time
- Check whether a manual finding is entrypoint-reachable
- Build an evidence packet for PoC work
- Review a single suspicious function discovered during manual audit
- Decide whether one issue should be promoted, deprioritized, or treated as
  part of a broader chain analysis

## When NOT to Use

- Multiple weak findings might compose into a stronger chain. Use a chain or
  composition workflow instead.
- The user wants a full audit. Use an audit or design-review workflow instead.
- The user wants remediation verification for a known finding. Use a
  remediation-review workflow instead.
- The target is a PR or branch diff. Use `graph-evolution` plus a differential
  review workflow.
- No concrete finding, function, file/line, or suspicious sink exists yet. Use
  discovery skills first.

## Rationalizations to Reject

| Rationalization | Why It Is Wrong | Required Action |
|---|---|---|
| "The scanner says high severity, so reachability is obvious" | Static findings need graph and code context before promotion | Bind the finding to a graph node and check entrypoint paths |
| "No entrypoint path means impossible" | It may mean parser, proxy, or dynamic dispatch limitations | Report the limitation separately from reachability |
| "An auth check appears on the path, so the issue is safe" | The check may enforce the wrong predicate or be bypassed by another path | Treat validation/auth as review targets, not proof |
| "One reachable path is enough for a PoC claim" | The path still needs attacker-controlled inputs and compatible preconditions | Separate graph reachability from exploitability |
| "This is probably a chain" | Single-finding triage stops at one candidate | Hand off related findings to a composition workflow |

## Workflow

```
Finding Triage Progress:
- [ ] Step 1: Normalize the candidate
- [ ] Step 2: Build or reuse the Trailmark graph
- [ ] Step 3: Bind the candidate to graph node(s)
- [ ] Step 4: Analyze reachability, taint, boundaries, and blast radius
- [ ] Step 5: Decide and emit the evidence packet
```

### Step 1: Normalize the Candidate

Accept file/line, function name, SARIF result, weAudit annotation, Markdown
finding excerpt, or a manual claim. Normalize it to:

- title
- source type
- file path and line range if present
- function or node hint
- suspected source, sink, or asset
- claimed impact

If there is no concrete code anchor, stop and ask for one.

For input handling details, see
[references/input-normalization.md](references/input-normalization.md).

### Shortcut for Steps 2–4 (file/line findings)

For file/line anchors, run the bundled script. It builds the graph and preanalysis once,
binds every finding, and prints the Graph Evidence fields as JSON: all matching nodes and the
primary one (the node containing the first line, then the narrowest), entrypoint paths with each entrypoint's trust level, `tainted` /
`entrypoint_reachable` / `privilege_boundary` / `high_blast_radius` membership, callers,
callees, downstream nodes, and the Trailmark version. Pass several findings in one call rather
than rebuilding the graph per finding.

```bash
uv run --no-project "{baseDir}/scripts/evidence_packet.py" --target <dir> --finding <file>:<line>[-<end>] --out evidence.json [--finding ...] [--language auto]
```

With `--out`, the complete packet stays on disk and stdout contains only delivery
counts. Read relevant finding records from that file. Without `--out`, stdout is
the full packet. Relative anchors may be target- or working-directory-relative;
ambiguous existing paths are rejected. Choose `--path-base target` or `cwd`, or
use an absolute path. Lines must be positive and ordered; column syntax is not
supported. Containers cover module/class/contract declarations when no callable
overlaps. Inspect `binding` (exact, nested or ambiguous), `binding_kind`, all
matches, unavailable signals and parser diagnostic limitations before judging the
result. A zero-node graph is a failure.
Batching shares graph work only: write a separate human verdict for each finding.

A finding that binds to nothing comes back with `bound_node: null` and the reason. The
script judges nothing: read the bound code and its callers, decide attacker control, and
treat an empty path list as a limitation to explain (dynamic dispatch, proxies, missing
entrypoint models), not as proof. Use the steps below for inputs the script does not take
(function names, SARIF via `audit-augmentation`) and for queries it does not run
(`connect_subgraphs`).

### Step 2: Build Or Reuse The Graph

Use the public `trailmark` skill workflow. Prefer an existing fresh exported
graph or `.trailmark/` artifact when present. Otherwise build a graph with
`language="auto"` or the target's explicit language list, then run
`engine.preanalysis()`.

Record the Trailmark version or feature probes used. Feature-gate Trailmark
0.4-only APIs with `hasattr()` or CLI help checks.

### Step 3: Bind The Candidate

Bind by file and line overlap first, then function name plus file. If several
nodes match, list every candidate and select the narrowest enclosing node as
primary. If no node matches, report a binding limitation instead of guessing.

SARIF and weAudit users should reuse the `audit-augmentation` workflow for
matching and then inspect the annotated node.

### Step 4: Analyze Graph Evidence

Run the query recipe in

From trailofbits/skills, CC-BY-SA-4.0.

More from trailofbits/skills