Methodology
ThreatDeck ingests public vulnerability databases, vendor advisories, research blogs and AI tool changelogs, triages them with language models, and publishes short, actionable cards. This page documents exactly how.
Priority score (0–100)
Deterministic and transparent. Weights below are rendered from the same code that scores cards.
| Factor | Points |
|---|---|
| Severity: Critical | 30 |
| Severity: High | 22 |
| Severity: Medium | 12 |
| Severity: Low | 5 |
| Severity: Info | 2 |
| Listed in CISA KEV | +30 |
| EPSS probability (scaled) | up to +15 |
| Public proof-of-concept | +8 |
| Affects AI systems / tooling | +12 |
| Card type: Act Now | +5 |
| Card type: AI Stack Alert | +5 |
| Card type: New Technique | +3 |
| Card type: Model Watch | +2 |
| Card type: Defender Kit | +1 |
| Card type: Policy | +0 |
| Card type: Signal | +0 |
The total is multiplied by 0.6 + 0.4 × confidence, so low-confidence items are pushed down, never up.
Exploit Clock
A rough estimate of how soon an issue will be weaponized. AI agents can turn public patch diffs and PoCs into working exploits quickly, so public PoCs weigh heavily.
- Exploited in the wild — Listed in CISA Known Exploited Vulnerabilities.
- Weaponization imminent — Public PoC on a high/critical issue, or EPSS ≥ 30%.
- Exploitation likely — Public PoC, or EPSS ≥ 5%.
- Exploitation unlikely (for now) — EPSS available but below 5%, no PoC.
- No signals yet — No exploitation signals published yet.
Guardrails
- Citation verification. Every card carries verbatim quotes from its sources. Quotes are machine-checked against the fetched source text; claims that cannot be matched are dropped.
- Held queue. Low-confidence cards and "Act Now" cards backed by a single unofficial source are held for human review before publishing.
- Unconfirmed label. Cards supported only by a single non-official source are marked "Unconfirmed".
- AI disclosure. Summaries are AI-assisted and labeled as such. Always read the original source before acting.
- Untrusted input. Source content is treated as data, never as instructions, and is escaped when rendered.
Dual-use policy
- We do not publish exploit code, weaponized payloads, or working prompt-injection strings.
- We do not publish AI-hallucinated package names that are still unregistered; those are shared privately with registry security teams.
- Detection rules and mitigations are published as AI drafts until verified.
- The tool radar lists publicly available offensive AI tools so defenders know what they face. Each carries defender notes; we link only to official repositories and never host payloads or step-by-step attack guidance.
Technique mapping
Stories and tools are mapped to the technique library: the OWASP Top 10 for LLM Applications (2025), the OWASP Top 10 for Agentic Applications (2026) and MITRE ATLAS. Mappings are AI-assisted and checked against the framework definitions.
Tool radar rings
Adopt, Trial, Assess and Hold are editorial ratings of maturity and usefulness for defenders, combined with repository signals (activity, releases, archival). Tools found automatically on GitHub stay in a separate "not yet reviewed" list until rated.
How AI is used
- What the model does. A language model (Anthropic Claude in production) decides whether an item is relevant, classifies it (domain, pillar, category, action) and writes the one-line summary, action note and brief. If the model is unavailable or over budget, deterministic templates are used instead.
- What the model does not decide. CISA KEV listings, CVSS and EPSS scores, advisory versions and minimum safe versions come from the source data. The card priority score above is computed by code, not by the model.
- What it sees. Public source text only, treated as untrusted data. No visitor data is ever sent to a model.
- Checks. Briefs must carry verbatim quotes that match the fetched source; unmatched claims are dropped. Outputs are screened for unsafe content (markup, scripts, shell one-liners, injection echoes) before publishing, and a fixed evaluation set is rerun whenever prompts or models change.
Source checking
- Official data first: CISA KEV, NVD, GitHub Security Advisories, OSV.dev and vendor advisories; then about 40 curated security outlets, research blogs and AI tool changelogs. The list and live feed health are on sources.
- The same event reported by several outlets is merged into one story that lists every outlet.
- Every story links to its original source; read it before acting. Briefs backed by a single non-official source are labelled "Unconfirmed".
- Playbooks, exposure sheets, detections and the vendor landscape cite their sources and show the date they were last reviewed; facts without a source are left out.
Corrections & retractions
Report an error to hello@threatdeck.net with the page URL and what is wrong. We check the claim against the source and reply.
- Corrected. A fixed item keeps its URL and gets a note "Corrected <date>" saying what changed. Feeds, API and MCP output carry the corrected version.
- Retracted. An item that should not have been published is removed from the site, feeds, API and MCP server.
- Takedowns (for example copyright) follow the terms.
API & MCP
Public and read-only. No key is required: anonymous callers get 60 requests per minute. A free
API key (Authorization: Bearer tdk_…) raises that to 600 per minute.
REST: https://threatdeck.net/api/v1/ (stories, cards, decks, AI dev tools and version checks, techniques, radar,
playbooks, landscape, detections, MCP Server Watch lookups, IOCs as JSON, CSV or STIX 2.1, search, personal RSS).
MCP (Streamable HTTP): https://threatdeck.net/mcp with 14 tools:
get_latest_deck, search_cards, get_card, list_ai_tools, check_ai_tool_version, get_top_stories, search, get_iocs, get_dashboard_panels, get_technique, search_tools, get_playbook, get_detections, check_mcp_server. Full reference on developers.