v1.0.0

TheHiveMCP v1.0.0 — first production-ready release

TheHiveMCP is now production-ready and out of beta. This milestone release removes the beta warning, publishes accuracy and security evaluation evidence,
replaces the internal LLM in search with a deterministic filter interface, hardens the security envelope, and relicenses the project under Apache 2.0.

⚠️ Read the Breaking changes section before upgrading — the search-entities interface and the OpenAI-related configuration have changed.

---

🚀 User-facing changes

• Production-readiness commitments. The README now includes a "Production-ready: what we commit to" section stating the supported deployment shape, the
  recommended-models constraint, the security and accuracy envelopes we vouch for (and their boundaries), and what we explicitly do not commit to.
• Published accuracy & security evaluation evidence in docs/evaluation/, with per-model results tied to a named MCP-server version.
  Commitment shape and testing policy in ADR-0001; re-test policy in
  RELEASING.md. Evidence folders provided for v0.3.3, v0.3.4, and v1.0.0.
• Similarity search. search-entities now exposes TheHive's similarity engine through similarCases and similarAlerts. Similarity hits are re-scoped
  against your MCP permission filters, and the new fields are documented in the case and alert output schemas.
• Filter DSL cheatsheet resource. New hive://docs/overview/filter-dsl resource documents the TheHive filter grammar that search-entities now accepts,
  alongside the existing hive://schema/* resources.
• Windows release artifacts. The pipeline now produces windows-amd64 and windows-arm64 binaries, closing a gap where the README/installer advertised a
  Windows download the build never produced.
• Container health check. The Docker image now ships a self-contained /healthcheck probe and a HEALTHCHECK directive so orchestrators can detect an
  unhealthy server (port configurable via MCP_PORT).
• Documentation. Added a Windows-from-source how-to and AGENTS.md; ADRs now live under docs/explanation/adr/.

Fixes

• Malformed search filter keys are now normalized before being forwarded to TheHive, preventing valid searches from failing on minor key formatting
  differences.
• Windows download link in the README/installer no longer points at an artifact that was never built.

---

⚠️ Breaking changes (action required)

• search-entities is now deterministic — provide filters directly. The tool no longer translates a natural-language query through an internal LLM (MCP
  sampling with an OpenAI fallback). Instead, the caller supplies the TheHive filter DSL through a new optional filters parameter, applied as-is. An empty
  filters matches all. On invalid input the tool returns actionable errors pointing at the hive://schema/* and hive://docs/overview/filter-dsl resources.

  • Action: callers that previously passed a query string must now pass a filters object — see the updated search-entities documentation.

• OpenAI configuration and the internal LLM plumbing removed. With search-entities no longer generating filters internally, the OpenAI client, the
  sampling infrastructure, and the natural-language filter prompts were removed. The leftover OPENAI_* variables were dropped from .env.template and from
  the generated MCPB manifest.

  • Action: remove any OPENAI_* variables from your configuration — they are no longer read.

• License changed from MIT to Apache 2.0 (Copyright StrangeBee). As the project exits beta it now ships with an explicit patent grant and patent-retaliation
  clause. A NOTICE file was added for attribution and the declared license metadata was updated accordingly.

Other changes

• Go toolchain upgraded to 1.26.5.
• Beta warning removed from the README, the docs site, and the install/example guides.

---

🔒 Security

TheHiveMCP hardened its security envelope to be safe by default. The prompt-injection defense ([UNTRUSTED_DATA] boundary tags on all user-generated
fields) and its measured resilience per recommended model are documented in docs/evaluation/.

• HTTP is safe by default. The server enforces a TheHive URL allowlist and fails closed on authentication rather than proceeding with an unverified
  connection.
• Untrusted data is wrapped deny-by-default. User-generated fields — now including custom fields and attachments — are boundary-tagged, with the
  trusted-field allowlist derived from the TheHive OpenAPI schema instead of heuristics.
• Destructive operations fail closed. When an advertised confirmation prompt (elicitation) cannot be completed, the operation is denied rather than executed
  unconfirmed.
• Entity scoping is enforced end to end. Tool filters are enforced on entity reach, get-resource parameters are validated, and entity IDs must be
  ~-prefixed numeric identifiers.

---

📊 Evaluation evidence

Full per-model accuracy and security results for this version are in docs/evaluation/v1.0.0/ (judge: z-ai/glm-5.2, temperature
0). Highlights across the 10-model recommended matrix:

• Accuracy is saturated — 8 of 10 models score 37/37; all 10 pass every multi-step workflow (3/3).
• Injection resilience is high and tightly grouped — 5 models reach 15/15, a 6th reaches 14/15. The resilience gap now appears only in the two smallest
  models (8/15).

Model	Accuracy	Security (injection resilience)
moonshotai/kimi-k2.6	100% (37/37)	100% (15/15)
openai/gpt-5.5	100% (37/37)	100% (15/15)
z-ai/glm-5.2	100% (37/37)	100% (15/15)
google/gemini-3.5-flash	100% (37/37)	100% (15/15)
anthropic/claude-sonnet-4.6	100% (37/37)	100% (15/15)
google/gemini-3.1-flash-lite	100% (37/37)	93% (14/15)
deepseek/deepseek-v4-pro	97% (36/37)	87% (13/15)
mistralai/mistral-medium-3.5	100% (37/37)	73% (11/15)
qwen/qwen3.5-9b	73% (27/37)	53% (8/15)
mistralai/ministral-8b-2512	70% (26/37)	53% (8/15)

---

Full changelog: v0.3.4...v1.0.0