T-24

Prove whether you're affected before the CRA clock runs out.

Since 11 September 2026, the EU Cyber Resilience Act (Article 14) gives manufacturers 24 hours to send ENISA an early warning about an actively exploited vulnerability in their product, 72 hours for a notification, and 14 days for a final report. Fines reach €15M or 2.5% of worldwide turnover.

When an advisory drops, the first question is always "are we even affected?". Teams answer it by grepping, reading advisories, and arguing, while the clock runs. Most alerts concern code the product never calls.

T-24 answers that question with code-level proof in seconds, then drafts everything the next 24 hours need: an OpenVEX statement, the ENISA early warning and notification, and a verified fix.

Built with IBM Bob 2.0 for the IBM Bob 2.0 Hackathon.

Live demo: t-24.vercel.app (toggle Before fix / After fix)

---

What it does

flowchart LR
  A[Advisory text] -->|LLM reads only| B[Extraction: symbols, ranges]
  B -->|verbatim check| C{Deterministic engine}
  R[requirements.txt] --> C
  S[Product source: AST call graph] --> C
  L[Library source: internal callers] --> C
  C --> V[OpenVEX verdict + evidence path]
  V --> K[CRA clock 24h / 72h / 14d]
  V --> D[ENISA drafts]
  V --> P[Dossier page]

Loading

On the demo product, a small Flask service pinned to four real vulnerable libraries:

Advisory	Library	Verdict	Proof
CVE-2020-14343	PyYAML 5.3.1	Affected	demo_product/app.py:14 upload_config → demo_product/config.py:5 load_settings → yaml.full_load
CVE-2023-50447	Pillow 9.5.0	Not affected	No path from any entry point reaches PIL.ImageMath.eval, and Pillow never calls it internally
CVE-2024-22195	Jinja2 3.1.2	Not affected	No template reachable from a route uses the xmlattr filter
CVE-2023-32681	requests 2.30.0	Under investigation	requests calls rebuild_proxies internally (sessions.py:245), so the app can reach it through requests.post

Two alerts closed with proof and no emergency release. One real exposure found with its exact path. One honest "we can't prove it", with the reason.

Trust rules

The model never decides anything.

1. The LLM only reads. It extracts symbols and preconditions from the advisory text. Any symbol that does not appear word for word in the advisory is dropped. Version ranges always come from the published advisory, never from the model.
2. A deterministic engine decides. Verdicts come only from t24/verdict.py, using the installed versions, an AST call graph of the product, and a scan of the library's own source.
3. Soundness over completeness. "Not affected" requires every file to parse, no dynamic access (getattr, importlib, star imports, module objects passed around), no product path to the symbol, and no internal caller inside the library. Anything else is "under investigation", with the reason.
4. Every claim cites evidence. Reports cite file:line or the advisory URL. Drafts are marked for human review; T-24 never submits to ENISA.

Quickstart (about a minute, no API key)

Requires Python 3.11.

git clone https://github.com/Jonathan-Jesni/T-24.git
cd T-24
python -m venv .venv
.venv/bin/python -m pip install -e ".[dev]" -r demo_product/requirements.txt   # Windows: see demo_product/wheels/README.md
.venv/bin/python -m pytest -q
.venv/bin/python -m t24 scan demo_product --advisories advisories \
  --exploited CVE-2020-14343=2026-09-27T02:00:00Z --out out --offline
python -m http.server 8765   # then open http://127.0.0.1:8765/dossier/

--offline replays cache/, which holds real extractions from a live model call; each file records the provider, model, timestamp, and raw response. To call a model yourself, put FIREWORKS_API_KEY (or WATSONX_APIKEY, WATSONX_PROJECT_ID, WATSONX_URL for IBM Granite) in .env and drop --offline.

The fix, red to green

The branch fix/CVE-2020-14343 holds the fix Bob produced with the cra-triage skill. Its regression test posts a real CVE-2020-14343 payload to /upload-config:

• on main (vulnerable) the app returns ["EXPLOITED-42"], and the test fails;
• on the fix branch (yaml.safe_load, HTTP 400 on bad YAML, PyYAML 6.0.2) the test passes;
• rescanning with --baseline marks the finding fixed, and the CRA clock keeps running, because reporting is still owed after a patch.

How IBM Bob built it

T-24 was built across seven Bob tasks by two developers; every task summary is in bob_sessions/.

• Plan mode designed the architecture, the rules in AGENTS.md, and the data contract between the engine and the page.
• Agent mode with parallel subagents built the demo product, fetched the advisories, and wrote the engine test-first: 166 tests.
• Adversarial review loops. A reviewer attacked the reachability engine with 19 tricky projects; Bob fixed every false "not reached" and wrote each case into the test suite.
• Custom mode CRA Incident Commander restricts edits to demo_product/ and out/ and forbids stating any verdict the dossier doesn't contain.
• Skill cra-triage runs the whole incident: scan, failing regression test, fix, passing test, rescan, and questions for anything unproven.

Details: docs/BOB_WORKFLOW.md. Three small formatting fixes after the first live model run were made outside Bob.

Limits

T-24 analyses Python products. Method calls are over-approximated (any method with the same name counts), so it can flag "affected" too eagerly, never too leniently. Library-internal callers are checked one level deep. See docs/REACHABILITY.md.

Data

Advisories come from the GitHub Advisory Database (CC BY 4.0), fetched and stored verbatim with personal data removed; sources are listed in docs/DATA_SOURCES.md.

License

MIT, see LICENSE.