Changelog
Unreleased - 2026-10-25
Added
- Persistent, operator-defined engagements with allowed targets, exclusions, authorization periods, and optional testing windows.
- Engagement IDs on active assessment tools, with target and time checks before execution. The checks apply to declared targets; they do not prove every secondary request made by an external tool remains in scope.
- Private on-disk evidence records with retrieval by engagement and record ID.
- Paged record and report lists; reports require explicit evidence IDs so older records cannot silently disappear from a client copy.
- Owner-specific metadata indexes so listing records and reports does not load every stored raw output across clients.
- Runtime diagnostics for tool availability and command limits, active Nmap scan listing, and an explicit succeeded/failed/cancelled outcome with run timing on evidence records.
- Analyst-authored findings and Markdown/JSON report retrieval with an evidence index, severity counts, assessment completeness section, optional affected assets/reproduction/retest details, and explicit wording when no findings have been verified.
- A Docker capture profile that opts in to host networking and the network capabilities required for packet capture on Linux.
- Versioned assessment and report prompts, assessment plans, context-quarantined Codex and Claude workers, provider profiles, worker-to-parent evidence import, and evidence-linked model report drafts with an analyst publication gate.
- Local CLI, pinned container, API-draft, and fixed SSH-draft profile types. Daybreak Blue and Red are optional access programs rather than product prerequisites;
local-agentkeeps engagement packets off hosted APIs. - OpenAI-compatible chat-completions profiles for loopback servers such as llama.cpp or Ollama and dedicated OpenRouter profiles. Both support report drafting; guarded assessment mode exposes only the plan tool intersection, enforces bounded calls and stop conditions, verifies the served model, and imports model-turn, denial, and run-counter audit receipts.
- Guided wordlist elicitation for authorized names, organization terms, dates, location fragments, and format rules, with private inputs and count/digest evidence instead of clue values.
- Hashcat 8-aware local credential checks: PCFG attack mode
4using the bundled grammar,hashRecipefor hash mode4000, and backend device IDs for hosts with supported accelerators. Runtime diagnostics expose the detected Hashcat version, and v8-only requests require an installed v8 binary. Native hosts still supply their own binary; FPGA bridge and hardware remain host-specific. - Opt-in
hostileModelProtectionfor duplicate action rejection, strict unknown-finish handling, and distrustful token accounting; it defaults off so intentional retests and ordinary provider variants do not cause alarms. - No-progress reflection, revisioned campaign state, ambiguous-interruption blocking, and
reconcileAgentAssessmentwith explicit operator confirmation and prior-worker scope revocation. - A real acceptance runner that requires explicit authorization and targets and can require separate tool-using assessment and report profiles.
- T3MP3ST design provenance and a complete Faxbot phased assessment dossier.
- A live OpenRouter draft, guarded-MCP run, and quarantined report-draft acceptance receipt using a hosted model that selected and invoked the real
generateWordlistMCP tool and saved an analyst-review report artifact.
Changed
- Stdio and loopback binding are the local defaults. Non-loopback HTTP requires bearer validation, issuer, audience, an allowed-origin list, and an HTTPS public base URL; the Docker HTTP profile always enables authentication and publishes to loopback by default.
- Authenticated HTTP requires separate read, write, and execute token scopes per MCP operation. Stateless HTTP creates a fresh SDK transport for each request, including successive calls by the same client.
- Authenticated HTTP scanner execution requires the separate
MCP_REMOTE_EXECUTION_ENABLED=trueopt-in. Host privilege auditing is available only through local stdio; hash cracking and packet capture are local stdio tools too. - External commands stop at the engagement deadline or when its owner closes the engagement, and honor MCP request cancellation. Failed and cancelled runs persist bounded audit evidence.
- Hashcat
deviceTypesnow selects OpenCL CPU/GPU classes with-Dinstead of device IDs with-d.backendDevicesselects IDs reported byhashcat -I; Hashcat 8 accepts but ignores the retainedworkloadProfilesetting. - The Dockerfile now builds and installs the signed Hashcat 8.0.0 source release with its PCFG rulesets and a CPU OpenCL runtime, replacing the older Kali package. GPU and FPGA passthrough remain host-specific.
- Scanner evidence now retains Nmap XML and bounded ffuf/Nuclei result files.
- Nuclei requires an operator-selected directory of reviewed, signed HTTP templates and host-wide authorization for its input targets. Its default invocation excludes network/OAST templates, redirects, and update checks.
- URL path exclusions reject ambiguous encoded paths, and enumeration tools reject path-restricted scopes they cannot honor.
- Docker runs as a non-root user and persists engagement data in a private bind mount. The image gives Nuclei templates to the runtime user and no longer grants passwordless sudo or privileged mode by default.
- Docker packaging now copies source before
npm ciso its prepare script can compile, and removes Go build caches from the runtime image. - Report scope text is taken from the registered engagement. Optional SoW notes in a report do not expand executable scope.
- Updated the MCP SDK, Inspector, Express, XML parser, Zod, and transitive dependencies to versions with zero findings in the current npm audit.
- npm packages now include only compiled code and documentation, so workspace scan journals, backups, and tests are excluded from published tarballs.
- Project licensing is now AGPL-3.0-or-later. Logging is always active and the credential-tool guides describe the enforced secret and evidence boundaries.
- Current Context7 provider documentation now drives OpenAI-compatible request behavior:
tool_choiceis omitted for Ollama compatibility, llama.cpp model paths are accepted, OpenRouter keys are pinned toopenrouter.ai, and OpenRouter requests require ZDR, data-collection denial, supported parameters, and no provider fallback.
Removed
- Legacy SSE transport, its Compose profile, and the
start:ssepackage script. - The unused
node-nmapdependency and tracked vendor files undernode_modules(the directory remains locally installed and Git-ignored). - Student/professional response modes,
setMode, conditional logging, the public synthetic agent profile, fake acceptance receipts, and the legacy tracked mode-change scan journal.
Security
- Issue 7: HTTP MCP requests always require bearer authentication, including on loopback; with authentication disabled,
/mcpreturns 401. Invalid tokens return 401 instead of 500. - John accepts only one approved
--wordlistand one--format. Hashcat rejects sessions, restore, loopback, file-backed masks and charsets, and positional inputs that begin with-. Input files on every transport must be owner-generated or underMCP_SHARED_WORDLIST_DIR, and tools receive the resolved path. runAgentAssessmentand model-backeddraftAssessmentReportover HTTP now requireMCP_REMOTE_EXECUTION_ENABLED=true, like scanner calls.- Every tool registers an explicit description.
Fixed
- Client Markdown reports keep each finding's Verified at, Retest status, CWE, and CVSS fields on separate lines. Researcher-written text and affected assets can no longer start lists, headings, horizontal rules, code blocks, or strikethrough. Leading indentation is removed and these markers appear as literal text. The completeness count of failed or cancelled runs only includes attached evidence.
[0.9.0] - 2026-02-18
Added
- New first-class tools:
trafficCapturehydraBruteforceprivEscAuditextractionSweep
- New report-admin tools:
listEngagementRecordsgetEngagementRecord
- SoW-aware reporting flow in
createClientReport:scopeMode=askuses MCP elicitation/user invocation- falls back to safe default template if declined/unavailable
- Invocation metadata output for the new tools (
invocation.clientId, session, scopes, request id when available). - Bundled MCP Inspector integration:
- dependency on
@modelcontextprotocol/inspector@^0.20.0 - new CLI entry path:
pentest-mcp inspector
- dependency on
- Command execution controls:
MCP_COMMAND_TIMEOUT_MSMCP_COMMAND_OUTPUT_LIMIT_BYTES
- Docker image support for modern recon dependencies (
httpx,ffuf,nuclei,subfinder) andhashcat. - Docker image support for
hydra,sqlmap, andtcpdump.
Changed
- Upgraded
@modelcontextprotocol/sdkto^1.26.0. - Streamable HTTP is now the canonical network transport (
MCP_TRANSPORT=http). - Node.js engine requirement updated to
>=22.7.5(Inspector compatibility). - Auth model modernized to bearer + OIDC/JWKS/introspection via:
MCP_AUTH_*MCP_OIDC_*
- Health endpoints now return transport/auth/deprecation metadata.
Deprecated
MCP_TRANSPORT=sseis now compatibility-only and emits deprecation warnings.- Legacy
MCP_OAUTH_*env aliases remain accepted temporarily but are deprecated.
Fixed
- Removed brittle dynamic transport/auth import fallbacks tied to outdated SDK structures.
- Corrected version reporting drift in runtime metadata.
- Hardened
httpxcommand resolution:- prefers
httpx-toolkit - validates fallback
httpxbinary - rejects accidental Python HTTPX CLI matches
- prefers
- Restored concrete Nmap XML parsing to populate
ScanDatahost/port/service details.
[0.5.0] - 2025-06-08
- Initial multi-transport release (
stdio,http,sse) with OAuth-oriented networking setup.