compliance is not security
|
A SOC 2 Type II report confirms controls existed during the audit period. It says nothing about whether those controls stop today's attacks.
"When Copy Fail is a 732-byte deterministic root with a public PoC, 'timely patching' is the wrong frame entirely."
three paths, one package
Pick the path that matches how you'll use this. No clone needed for most users — one npx command and your assistant has access to every skill, every index, every catalog.
AI consumer
Point your AI assistant at the installed package. It collapses the seven-phase contract into discover → brief → run for the conversational flow, or ai-run --stream for JSONL pipe.
$ npx @blamejs/exceptd-skills brief secrets
$ npx @blamejs/exceptd-skills run secrets --evidence -
# or one streaming pipe:
$ npx @blamejs/exceptd-skills ai-run secrets
operator
Refresh CVE/RFC data, run currency checks, generate reports. GHSA is a refresh source (covers npm + PyPI + Maven + Go) so fresh disclosures land within hours, not the days KEV/NVD take. refresh --advisory <CVE-ID> seeds a single draft entry on demand; refresh --curate <CVE-ID> surfaces editorial questions + ranked candidates (ATLAS / ATT&CK / CWE / framework gaps) for a human reviewer or AI assistant. Air-gap workflow unchanged.
$ npx @blamejs/exceptd-skills refresh --curate CVE-2026-45321
$ npx @blamejs/exceptd-skills refresh --network
$ npx @blamejs/exceptd-skills ci --scope code --max-rwep 70
maintainer
Add skills, update catalogs, cut releases via tag push. Ed25519 signing, 18-gate predeploy, and npm provenance baked into the publish flow.
$ npm run bootstrap
$ npm run predeploy
what ships in the box
Twelve things every other GRC tool either doesn't have or implements as a checkbox.
seven-phase playbook contract
govern → direct → look → detect → analyze → validate → close. exceptd owns the knowledge + GRC layer; your AI assistant owns artifact collection. The CLI is brief (info doc) + run (execution) + ai-run --stream (JSONL pipe). Closure ships a signed CSAF-2.0 / SARIF / OpenVEX / markdown bundle, computed jurisdiction deadlines, and auditor-ready policy-exception language.
nightly upstream refresh + minutes-old disclosures
Automated GitHub Actions job pulls KEV, NVD, EPSS, IETF, MITRE, and the GitHub Advisory Database every 24h. KEV listing takes days; NVD ~10 days; GHSA fires within hours and covers npm + PyPI + Maven + Go + NuGet + …. 15 primary-source feeds (advisory venues + vendor research blogs + tech-press + a GitHub public-events tracker for named researcher handles — first registered: Nightmare-Eclipse / Chaotic Eclipse, anchoring NEW-CTRL-073). A complementary CVE regression watcher (NEW-CTRL-074, lib/cve-regression-watcher.js) cross-checks poller diffs for historical-CVE references that may indicate silent vendor regression — the class anchored by MiniPlasma re-breaking CVE-2020-17103 in 2026. The CVE catalog itself now spans 1,225 entries — anchored by a CISA KEV bulk-intake (68 → 312 in a single pass) covering all dateAdded >= 2024-01-01 actively-exploited entries. For "I want this CVE today, not tomorrow": exceptd refresh --advisory <ID> --apply.
RWEP scoring
Real-World Exploit Priority. CVSS + KEV + EPSS + PoC availability + AI-assist + live-patch + blast radius. Calibrated to mid-2026 weaponization timelines.
35 jurisdictions
EU AI Act, NIS2, DORA, GDPR · UK CAF · AU Essential 8, ISM · SG MAS TRM · JP NISC · IN CERT-In · CA OSFI · and more. Breach clocks normalized.
Ed25519-signed skills
Every skill body signed against a per-repo keypair. Tamper-evident at runtime. Dual SHA-256 + SHA3-512 public-key fingerprints for out-of-band pinning. npm audit signatures verifies npm provenance.
17-gate release hygiene
Internal release gate run by maintainers before every npm publish: Ed25519 verify · tests · catalog schemas · manifest snapshot · skill lint · watchlist · catalog-meta · SBOM currency · pre-computed indexes · vendor tree · publish tarball shape · shipped-tarball signature round-trip · diff coverage · playbook validation · test-count baseline · catalog-gap budget · version-tag drift. Distinct from the operator-facing exceptd ci verb (which runs playbooks with --max-rwep thresholds against your repo).
multi-format evidence bundles
Every run emits a signed CSAF 2.0 envelope plus optional SARIF 2.1.0 (GitHub Code Scanning, VS Code SARIF Viewer, Azure DevOps), OpenVEX 0.2.0 (sigstore / in-toto / GUAC), and markdown digest. Matched CVEs + indicators that fired + framework gaps all surface as first-class findings — posture-only playbooks still emit a meaningful bundle.
AI-first conversational contract
ai-run --stream emits JSONL phase events; the AI submits evidence back on stdin; phases 4-7 stream as they complete. One pipe, no file handoff. ask "<question>" routes plain English through synonyms to the right playbook. discover auto-detects cwd context and recommends.
terminal-first output, no jq required
Default ci output is a one-screen digest: verdict line + per-playbook table (id · verdict · rwep · evidence · finding) + deduped session warnings + scope rules + framework-gap rollup + pending jurisdiction obligations + per-verdict Next steps block with concrete exceptd lint <pb> - / --format csaf-2.0 commands per detected playbook. run result envelopes hoist verdict · rwep_score · top_finding · evidence_completeness · attestation_path to the top level. attest verify / attest diff answer their one-line questions at the terminal. --json / --pretty reach the structured envelope when automating.
supply-chain capability + provenance triage
Reads your dependency tree through an eight-class capability lens — network · filesystem · shell · env · eval · install-script · telemetry · native-binary — independent of any known CVE. Surfaces likely capability-creep when a dependency gains a class it never exercised in the prior version, and the install-script + shell + network + env credential-harvest shape even at zero matched CVEs. Screens for typosquat / homoglyph names, obfuscated-only source, and dependency-confusion resolution (internal scope served from the public registry, inflated-version public squats). CVE-reachability demotion separates runtime-present from build-only matches without ever suppressing a tree-presence match. Publisher-identity-change detection surfaces a likely silent ownership-handoff precursor that capability-creep can't see. A 0–100 package_confidence trust score reports alongside RWEP and never feeds or replaces it — RWEP scores known-vuln urgency, package-confidence scores provenance.
offline-first, zero runtime dependencies
The full CVE / MAL catalog, every skill, and all 33 playbooks ship inside the package — exceptd reads them from local files with zero runtime dependencies. No network call is required to run a scan and no telemetry is emitted. Every playbook carries an air-gap mode (--air-gap) that substitutes bundled snapshots for network-touching sources, so analysis runs unchanged on a disconnected or classified host.
33 attack-class playbooks
One per attack class — kernel LPE, MCP supply chain, container escape, dependency confusion, credential stores, post-quantum exposure, CI/CD compromise, ransomware, identity / SSO, LLM tool-use exfiltration, and more. Each runs the seven-phase loop and loads only the skills it needs. brief --all lists the set; run <playbook> walks it with your evidence.
from upstream to your agent
discover recommends playbooks for your cwd. brief shows you what one playbook checks. run walks the seven phases with your evidence. attest + doctor close the audit loop.
discover
Scans your cwd (.git/, lockfiles, Dockerfile, k8s manifests) + host platform, then recommends the relevant playbooks with reasons.
exceptd discoverbrief
One info doc — jurisdictions + threat context + RWEP threshold + preconditions + artifacts + indicators. Collapses phases 1-3.
exceptd brief secretsrun
Phases 4-7. Submit evidence as flat observations + verdict JSON. CSAF / SARIF / OpenVEX bundles + Ed25519-signed attestation.
exceptd run secrets --evidence -attest
Auditor-facing operations: list · show · export · verify · diff. Stored under ~/.exceptd/ with repo-tag.
exceptd attest list51 skills. all current.
showing 12 · last_threat_review 2026-05-15Skills are read-only knowledge units. To execute, you run a playbook — there are 33 of them, each loading the relevant skills. Use exceptd brief --all for the playbook list, exceptd run <playbook> to execute, and exceptd skill <name> to read a skill.
researcher
Drop in a CVE ID, ATLAS TTP, advisory, or incident narrative. Gets routed to the right skill automatically.
kernel-lpe-triage
Copy Fail (732-byte root, deterministic) and Dirty Frag (IPsec/RxRPC chain). Live-patch vs. reboot decision tree.
ai-attack-surface
Prompt injection RCE, MCP supply chain, RAG exfiltration, AI-assisted exploit dev. ATLAS v2026.07 mapped.
compliance-theater
7 documented theater patterns with specific detection tests. SOC 2 pass ≠ Copy Fail immunity.
global-grc
EU AI Act, NIS2, DORA, UK CAF, AU Essential 8, SG MAS, JP NISC, IN CERT-In. 35 jurisdictions.
exploit-scoring
Real-World Exploit Priority beyond CVSS. KEV, PoC, AI weaponization, active exploitation, patch delta.
mcp-agent-trust
MCP trust boundary failures, unsigned manifests, tool allowlisting gaps. CVE-2026-30615 zero-interaction RCE.
pqc-first
Post-quantum cryptography migration. NIST PQC standards, harvest-now-decrypt-later exposure, migration roadmap. exceptd discover probes 22 PQC algorithms across NIST finalists, Round-4, signature on-ramp, stateful, and composite families.
zeroday-gap-learn
Zero-day → attack vector → control gap → framework gap → new control requirement. Feeds back into all skills.
supply-chain-integrity
Eight-class capability taxonomy, capability-creep + credential-harvest shape, typosquat / homoglyph / obfuscation / dependency-confusion screens, CVE-reachability demotion, publisher-identity-change, package_confidence trust score.
identity-assurance
Authenticator, proofing, and federation assurance levels (AAL / IAL / FAL) per NIST SP 800-63. WebAuthn / FIDO2 passkeys, phishing-resistant MFA, and OIDC / SAML / SSO trust-boundary failures.
coordinated-vuln-disclosure
Run a disclosure program to ISO 29147 / 30111 — security.txt intake, CSAF 2.0 advisory authoring, VEX statements, and embargo / timeline coordination across affected vendors.
curated chains for common workflows
8 pre-built sequences in data/_indexes/recipes.json. Load one when planning a multi-skill session.
Drop a CVE ID. Get RWEP score, framework gaps it exposes, and a learning-loop entry.
Full AI/ML attack surface assessment before a red-team engagement.
Find theater gaps, document framework insufficiencies, generate auditor-ready exceptions.
Threat-led penetration testing scoped to DORA Article 26 requirements.
18 pre-computed indexes
AI consumers answer cross-reference questions without scanning every skill and catalog. Regenerate with exceptd refresh --indexes-only. Incremental rebuilds via --changed.
100-word skill abstracts
What to load when planning a multi-skill workflow. 12 KB total for all 51 skills.
defense-in-depth ladders
Canonical DiD ladders per attack class: prompt injection, kernel LPE, AI-as-C2, ransomware, supply chain.
compliance theater patterns
7 structured records: claim, audit evidence, reality, fast detection test, controls implicated.
breach notification matrix
29 jurisdictions × obligation × hours. Patch SLA and notification deadlines normalized.
CVE/CWE crosswalks
Per CVE and CWE: which skills cite it, which framework gaps surface, which D3FEND countermeasures apply.
context cost per skill
Approximate token cost per skill and per section. Load before composing multi-skill sessions.
what every other tool gives you
why this exists
Framework lag is measured in months. MITRE ATLAS v2026.07 is the most current AI threat framework available. It still lags real exploitation by 3–6 months. NIST AI RMF lags by years. ISO 27001:2022 has no AI-specific controls. Every skill explicitly flags where framework coverage ends and real attacker capability begins.
AI changed the exploit-development timeline. Copy Fail was discovered by an AI system in approximately one hour. 41% of 2025 zero-days involved AI-assisted reverse engineering on the attacker side. Risk scoring must reflect this — CVSS does not.
Every org has a compliance theater problem. The question is not whether paper controls map to audit requirements. The question is whether those controls would actually detect or prevent an attack. These skills answer the second question.
stop running on stale intel
Open source. Apache 2.0. Works with Claude Code, Cursor, Codex CLI, Gemini CLI, Windsurf, and any AGENTS.md-aware assistant.