The existing voters answer "is this finding real?", which invites judging the
whole narrative at once — and that is how a proven missing control got deleted
for having an overstated headline. One lever, one verdict, observation gone.
The Evidence Prosecutor has a narrower brief and four questions in order: what
exactly was observed, which sentences go beyond that, what would have to be
observed for the claimed impact to be factual, and — the one that decides
retain-versus-reject — would anything security-relevant remain if the
unsupported sentences were removed.
It cannot pass sentence. There is no verdict field in its contract (a test
asserts the prompt never offers one), and apply() can only narrow: the minimal
supported statement replaces the mechanic, the asserted impact is demoted to
potential with the conditions that would make it factual attached. Nothing it
returns can raise a severity, add an impact, or drop a finding.
A self-contradicting verdict — "nothing survives" alongside a minimal claim —
is detected and the reading that keeps the observation wins: the claim is a
concrete artifact, the boolean is an opinion about it.
Findings from before the claim contract get a ledger reconstructed from what
they recorded (structured evidence where present, quoted evidence lines
otherwise), so the back catalogue is judged rather than silently skipped.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two halves of the same problem — proving what actually happened rather than
what a response suggested.
browser.rs — a payload echoed into HTML is reflection; it is XSS only when a
browser parses that response and runs it. The gap between the two is where most
XSS false positives live: the value lands in an attribute that is never
evaluated, inside a <textarea>, HTML-encoded on the way out, or blocked by CSP.
All four look identical to a string match on the body.
So a real Chromium loads the URL and reports whether a marker THE HARNESS CHOSE
came back through a channel only executing code can reach: a dialog message, a
document.title assignment, a window global. A console line is watched too but
never treated as decisive on its own — a page can log the value it reflected
without ever running it. Payloads are self-reporting rather than generic
(alert(1) proves nothing attributable) and cover the contexts a reflected value
lands in: raw HTML, attribute break-out, event handler, URL, raw-text element,
template expression. When the marker is reflected but does not execute, that is
recorded as a note: it tells the operator the input reaches the response and
the context is what stopped it.
Missing node or playwright yields available:false and confirms nothing. A
missing tool must never read as a missing vulnerability — or as a present one.
replay.rs — the engagement recorded 25 accepted POSTs and concluded "reset
email flooding". Those are facts at different layers and the pipeline could not
say so. observe_effects() now records three:
request_effect the response: status, headers, latency, size
application_effect a read-back showing state actually changed
external_effect something left the building (mail, webhook, job)
Without a verification request the application layer is reported as unexamined
rather than inferred from a 200 — APIs accept and ignore writes routinely. The
external layer is honestly reported as unobserved until the harness owns a
mailbox or callback listener. deepest_observed() gives the ceiling an impact
claim may be built on, which is exactly the line the agent crossed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The Arena engagement rejected "no rate limiting on the password-reset flow"
outright. The agent had proved 25 requests accepted with no 429, no
Retry-After, no RateLimit-* — and then titled it "reset-email flooding". The
voter judged the claimed impact unproven and discarded everything, so a real
missing control never reached the report.
That was structural, not a bad call. The judge got a prose paragraph and one
accept/reject lever, while agents reliably walk: control absent -> abuse
possible -> impact plausible -> impact written as fact. The chain has to break
at step two, and that needs the finding to arrive as separable claims.
claims.rs adds:
- An evidence ledger (E01, E02, …) so a verdict is auditable: "supported by
E01-E27" is checkable, "the evidence looks convincing" is not. A claim citing
an id that was never recorded is REJECT_INVALID_EVIDENCE — worse than citing
nothing, because it looks supported.
- Mechanic and impact as separate claims, each with its own citations. An
asserted status never outruns its evidence: a model may downgrade itself and
can never upgrade past what it cited.
- Six structured decisions instead of accept/reject. DOWNGRADE_UNPROVEN_IMPACT
and DOWNGRADE_SCOPE_LIMITATION cannot discard — that is enforced by
Decision::discards(), not by an instruction a model could reinterpret.
- Impact preconditions: "email flooding" needs account_exists +
account_confirmed + email_delivery_observed. 25 accepted requests prove
throttling was not observed; they do not prove mail was delivered. The
difference is now computed, not argued.
- A rewriter, because lowering severity is not enough: a report headed
"Password Reset Email Flooding" still asserts flooding whatever number sits
beside it. The title is rebuilt from the mechanic, the impact prose becomes
Observed / Not demonstrated / Potential impact, and the claimed consequence
survives only as clearly labelled potential.
- still_security_relevant(): strip the unproven impact and ask whether anything
remains. "The reset endpoint has no observable rate limiting" does; "the
application responded" does not. That decides retain-vs-reject.
Wired into the pipeline ahead of the voters, and the voter can no longer delete
a finding that arrived with claims — it can only mark the narrative rejected
while the mechanic stands.
A test caught an inverted comparison in the severity cap that silently left a
High finding at High: a cap must lower and never raise.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Four pieces that together answer "may this action happen, under whose
authority, and can we prove afterwards what we did".
policy.rs — effective_risk per action, exactly as specified:
(action_risk + asset_criticality + protocol_risk + privilege_level
+ blast_radius) × environment_multiplier
Every term is named and kept on the result, so the number can be explained
rather than argued with. Three policies sit on it: SafetyPolicy (ceilings,
approval thresholds, hard prohibitions), ReasoningPolicy (baseline before
payload, bounded hypotheses, evidence before escalation, explicit stop
conditions) and ProofOfImpactPolicy (what a severity must carry before it may
be that severity).
OT/ICS/SCADA is treated as its own regime, not web testing on odd ports.
Industrial protocols authenticate nothing — a Modbus write is the protocol
working as intended, addressed to a device that may be holding a valve — and
scanners crash PLCs by sending unexpected data at line rate. So the OT profile
blocks writes, disruptive actions, fuzzing and exploit payloads outright, caps
the rate at ~1 req/s, and refuses the function codes that stop a CPU (Modbus
5/6/8/15/16/22/23/43, S7 start/stop, DNP3 restart/stop). Safety instrumented
systems are off limits in every profile.
A test caught a calibration error worth keeping: a plain READ of a critical PLC
scores 3.6 on this formula, so the obvious tight ceiling would have refused
exactly the observation OT findings come from. In an industrial environment it
is the KIND of action that is forbidden, not the arithmetic — the ceiling
catches extremes and the low approval threshold makes anything past trivial
observation a human's decision.
capability.rs — HMAC-signed grants: who authorized what, against which hosts,
in which environment, until when. The harness verifies the signature before
reading a single claim (a well-formed token from the wrong key must never get
to influence what the harness believes), refuses expired and not-yet-valid
tokens, and treats the grant as a CEILING: constrain() intersects it with local
configuration, so config can narrow authorization and never widen it. Tokens
carry no secrets — the payload is readable by anyone holding it.
audit.rs — one structured record per action, in the specified shape (timestamp,
agent, hypothesis, action, target, policy_decision, operator, tool, result,
evidence_hash, capability_token). Two things make it worth having: it is
hash-chained, so removing or editing an entry breaks every hash that follows
and verify() says which one; and it records REFUSALS, because a trail
containing only what happened cannot demonstrate restraint. Only the grant's
id is recorded, never the token — the trail gets shared.
Hard kill conditions end a run outright: target unresponsive after our traffic,
sustained 5xx, out-of-scope request, forbidden industrial function code, safety
system addressed, capability expired mid-run, repeated policy violations,
budget exhausted, operator stop. Failures BEFORE the target ever answered do
not count — nothing listening is not the same as knocked over. The OT switch
trips far sooner: a PLC missing two requests already warrants stopping.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Validators decide from recorded artifacts, and the weakest link was who
recorded them: "the payload returned a 500" is still an agent's account of what
happened. Replay produces the part that matters most in practice —
reproducibility — by sending the request again through the harness's own
client, with the scope guard in front of it.
Three properties it is built around:
- every request passes ScopePolicy::check_request before a socket is opened, so
replay cannot be the thing that wanders off-scope while verifying a finding;
- it never mutates: a finding proven with DELETE is not re-proven by deleting
the record again, so non-idempotent verbs are refused and repeats of them are
refused outright;
- bodies are truncated at 96KB and SAY they were truncated — a silently clipped
body makes a length differential meaningless.
enrich() fills in repeats and re-measures a recorded baseline (comparing a
fresh attack against an hour-old baseline attributes ordinary drift to the
payload). It deliberately does NOT synthesize a baseline from an attack
request: removing "the payload" from an arbitrary URL is guesswork, and a
guessed baseline would silently decide the verdict.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two gaps this closes, both found by reading what the code actually did.
Scope was never enforced
------------------------
`out_of_scope` was rendered into the prompt as "HARD CONSTRAINT — do NOT test…"
and nothing checked it. That is a request to a model, not a control: an agent
that decided a discovered subdomain was interesting, or that followed a
redirect off-target, was free to act and the operator found out by reading the
report.
scope.rs adds a guard in code:
- Hard scope: allowlist of hosts, *.wildcards, IPv4 CIDRs, URL prefixes, with
exclusions that always win. Defaults to the engagement's own target, so
discovery cannot widen authorization — finding a host is not permission to
attack it. An unconfigured policy is closed, not open.
- Soft scope: observe-only zones, destructive verbs (off by default), an
account-creation cap, a rate guard that warns rather than silently dropping
requests (a dropped request reads as "target unreachable"), and payload
classes refused even in scope because they damage the target instead of
demonstrating a bug.
- Enforced at the harness's own chokepoint (probe) and as a post-run audit:
findings proven against an unauthorized host are withheld from the report and
written to out-of-scope-findings.json as an incident to disclose, because
shipping one would launder the mistake.
- REPL: /inscope, /observe, /guardrail, /policy; /scope-out now promotes
host-shaped entries into enforced rules immediately, and says plainly when an
entry is prose the guard cannot enforce.
Validation was models checking models
-------------------------------------
N-model voting plus an adversarial refute pass share the failure mode of the
thing they check — agreement is not evidence, and a confident hallucination
survives a vote by being confident. grounding.rs helps but matches keywords
("http/", "status", "alert(") and cannot tell a real response from a plausible
transcript of one.
validation.rs asks a different question — does the recorded evidence
demonstrate THIS class? — with per-CWE rules and no model in the loop:
SQLi baseline/attack difference that reproduces >= 2x
XSS a browser executed a harness-chosen marker; reflection is not proof
IDOR identity B reads A's resource AND the body matches (a 200 returning a
login page is rejected, which is the classic false positive)
SSRF controlled callback or canary retrieval
LFI controlled marker or a file signature the baseline lacked
RCE a unique nonce in output/callback; reflected input is rejected
Absent evidence is never a pass, and a class with no rule is never
auto-confirmed. NEUROSPLOIT_VALIDATION=advisory (default) rejects
contradictions without demoting voted findings for missing artifacts;
enforcing makes the verdict the status. The evidence contract is injected into
exploit prompts so agents collect the artifacts while they still hold the
target.
Finding gains evidence_data so agents can emit structured artifacts alongside
the finding JSON.
Two bugs the tests caught while writing this: the scope guard treated a SAST
`src/auth.rs:42` endpoint as a host and quarantined valid source findings, and
two canaries minted in the same clock tick came out identical — a marker that
repeats would let a stale token vouch for a new finding.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Backend
-------
- knowledge_graph.rs — the durable structure under attack_graph's per-run view:
typed entities (asset/endpoint/weakness/technique/finding/account/credential/
impact) joined by typed, weighted, provenance-carrying edges, accumulated
across runs in .neurosploit/graph.json plus a per-run copy the report and web
console can draw. Answers what a finding list can't: ranked attack paths, and
the frontier of entities observed but never proven — where chaining should
look next. Agents only sometimes fill chains_from, so progression is also
inferred between adjacent kill-chain stages; those edges are marked inferred,
weighted lower, and drawn dashed, because presenting a hypothesis as evidence
is the graph lying about itself. Secrets stay in the vault, never the graph.
- memory.rs — four tiers scoped by lifetime, not importance: working (one run),
engagement (one target), technique (one agent/CWE), reusable (generalized).
Promotion is evidence-gated and needs independent evidence at each step: a
claim repeated within a run becomes engagement knowledge; one confirmed
across runs becomes technique knowledge; one that held on two DIFFERENT
targets is generalized into a reusable lesson with host-specific tokens
stripped. Nothing is promoted on a single observation, which is exactly what
a hallucination looks like. Recall is scored (overlap × past success ×
recency) and injected into recon/exploit prompts as leads to verify. Recalled
memos are credited only when the run they informed actually found something.
- rectify.rs — a mistyped command cost a full round trip through /help, at the
worst possible moment during a live run. Accepted-as-typed wins over
everything (so the /url alias is never "corrected" to /ua), then unique
prefix, then Damerau-Levenshtein with a length-scaled budget, and a tie is
reported rather than resolved. Arguments too: a bare host gets its scheme, an
out-of-range count is clamped with a note instead of silently reverting, a
near-miss model id is matched against the live catalog.
- pool.rs — when every configured model is exhausted or its token is dead, try
whatever else this machine can actually reach (an installed CLI subscription,
or a provider whose key is in the environment) before parking. A run that
stops on a box with three other usable backends stopped for no reason.
- repl.rs — /memory, /forget, /graph; a recovered run resumes by itself where
nobody is watching (piped stdin — the web console — or NEUROSPLOIT_AUTO_RESUME),
since a `/continue` prompt there waits forever.
Web
---
- Attack path: the stage list was seven hardcoded values, so findings the
harness staged outside it were silently dropped — 5 of 27 on a real run.
Rewritten against the harness's own stage list with unknown stages kept,
two-line labels (every node used to read "SQL Injection Authent…"), stage
column headers, pan/zoom/fit, path highlighting, severity filter, and the
run's graph.json used when present.
- Dashboard: coverage, findings by severity, top weaknesses, and annualized
loss exposure via FAIR — frequency from exploitability × validation
confidence, magnitude from assumptions shown on screen and editable, reported
as a range. The posture score saturates instead of subtracting, so it keeps
discriminating past the first critical.
- Run history groups into one folder per target with a filter, instead of one
flat list that grows forever.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BvdGy9XtVWSdXDTa3FFLJv
- Add 12 technique/scenario LLM red-team agents (AI category 18 → 30, total 429):
jailbreaks — AdvPrefix, PAIR, TAP, Crescendo, many-shot, persona/DAN,
encoding/obfuscation, refusal-suppression; prompt-injection scenarios — direct,
indirect (RAG/web/email/tool output), goal hijacking, tool/function-call abuse,
system-prompt/secret exfiltration. Each runs an attacker→LLM-judge loop
(baseline refusal → technique across variants → verdict), proving the bypass
with a benign, redacted receipt. Generated by scripts/build_llm_redteam_v365.py.
- Add REDTEAM_DOCTRINE and inject it into run_ai so every AI test follows the
baseline→technique→judge method across scenarios.
- Models: add Claude Opus 5 and Sonnet 5 (Anthropic) and a new Moonshot AI (Kimi)
provider with Kimi K3/K2 (moonshot:kimi-k3, MOONSHOT_API_KEY) — 15 providers.
- Docs: README/TUTORIAL/RELEASE — new AI/LLM red-team engagement mode + section,
model/env-key tables, agent-library counts (429), badges.
Also includes the v3.6.4 grounding fix (#33) landing on main.
The grounding gate ran in empirical mode for every engagement, demoting
white-box (and skills/n8n audit) findings that had passed the n-model vote
because a file:line code citation isn't raw tool output. Grounding is now
mode-aware:
- Symbolic (white-box SAST / skills): a file:line reference into the reviewed
source, or a quote of code present in it, is the receipt — no live target.
- Empirical (black-box / host / AI): evidence must resemble tool output (as before).
- Either (grey-box): a source citation OR a tool receipt grounds a finding.
The symbolic check runs against the reviewed source corpus (not the transcript)
and falls back to a structural file:line + quote check when the corpus is
unavailable. Adds unit tests incl. a regression test for #33.
- /continue (and /resume) now relaunch a recovered interrupted run on the same
target, carrying its findings forward and steering agents to widen coverage /
chain from them instead of re-reporting. Offer shown at launch; a fresh /run
supersedes it. Findings merge (dedup by title+endpoint) across both runs.
- Opening /results, /finding or /report while a run streams no longer corrupts
the terminal: live background output is paused for the picker (still captured
in /logs) and restored on exit, so Ctrl-C in a picker can't take the process
down mid-run.
- Drive `codex exec --json` and parse its JSONL event stream into the same
categorized live feed as Claude Code (exec/edit/tool/net/tokens), so recon and
exploitation are visible as each command runs instead of a silent black box.
- Fix the activity feed to keep per-agent tool events (commands, network, files,
findings) and only filter model reasoning + token telemetry, so /logs shows the
real command trail and /status 'last:' is a true sign-of-life.
- Surface failed internal commands as 'exec: (exit N)'; keep Codex auth/rate
detection from stderr.
Added the OpenAI GPT-5.6 line to the provider pool: gpt-5.6-sol (frontier/default),
gpt-5.6-terra (balanced), gpt-5.6-luna (fast/affordable). Version 3.6.0 -> 3.6.1.
New harness::probe runs a real request/response analysis of the target BEFORE
the model recon and injects the observed facts into recon, so agent-selection
and exploitation decisions are grounded in evidence (robust even when model
recon is weak):
- status & redirect, Server/X-Powered-By/content-type, 6 security headers,
cookie flags (HttpOnly/Secure/SameSite), CORS reflection test (arbitrary
Origin + credentials), tech fingerprint, linked scripts, form count, a 404
baseline for soft-404 differentials, and high-signal paths (/robots.txt,
/.git/config, /.env, /sitemap.xml, /.well-known/security.txt).
- Best-effort (never fatal — degrades to a note on network failure), honors the
identifying User-Agent and the Burp/ZAP proxy. Wired into black-box run() and
greybox recon. A one-line probe summary streams to the live feed.
New harness module `integrations` (+ app commands) wiring NeuroSploit into the
SDLC. Config persists per-project to .neurosploit/integrations.json; secrets are
NEVER stored — only the env-var name is saved, values read from the environment.
GitHub:
- private-repo clone (token injected into the clone URL for whitebox/greybox/tui)
- `neurosploit pr <owner/repo> <n>`: clone the PR head (refs/pull/N/head),
white-box review, optional `--comment` (PR summary) and `--jira` (cards)
- `neurosploit watch <owner/repo> --branch --interval`: re-review on each new commit
GitLab:
- private-repo clone (oauth2 token) for whitebox/greybox (gitlab.com or self-hosted)
Jira:
- `--jira` on any engagement opens one card per finding (REST /issue, basic auth)
Control:
- `/integrations` (REPL): show · enable/disable · setup jira|gitlab|github
- `neurosploit integrations [show|enable|disable] [github|gitlab|jira]` (CLI)
Docs: README "Integrations" section + new TUTORIAL-INTEGRATION.md (per-tool setup,
scopes, recipes, troubleshooting). Version bumped 3.5.2 → 3.5.3.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Distilled from reviewing real AI-pentest output that kept stopping at "exposed"
instead of "exploited". Pure-additive, back-compatible.
Behavior (injected into black/grey/chain exploit prompts via DEPTH_DOCTRINE):
- Exposed → exploited: any info-disclosure / exposed service/WSDL / leaked
credential|token / reachable dev host MUST be used before it's a finding;
otherwise it's a lead, not a confirmed High/Critical.
- Chain across modules: reuse obtained session/JWT/cookie/credential and pivot
to IDOR/privesc/exfil; report the chain, not isolated parts.
- Decode & fingerprint → CVE; audit tokens (alg-confusion/none/kid/JWKS, weak
HS256 secret cracking, lifecycle).
Deterministic post-pass (new crates/harness/src/hygiene.rs, wired into finish()):
- calibrate severity to PROVEN impact — unproven High/Critical (hedged, no
payload, thin evidence) capped to Medium and re-titled "(potential)";
- depth_audit — flag exposures on a host with no real exploit;
- hygiene_summary — advise consolidating hygiene classes repeated across assets.
Unit tests cover calibration + depth audit.
5 new doctrine meta-agents (scripts/build_methodology_v352.py → agents_md/meta/):
exploit_depth_doctrine, finding_chainer, artifact_decoder, token_auditor,
report_calibrator (meta 17→22, total 343→348).
Version bumped 3.5.1 → 3.5.2 across crates/app/installers/docs; RELEASE/README
updated.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Partial observability is now first-class:
- belief.rs — property-graph world model; nodes (host/service/vuln/exploit/cred)
carry a probability, not a boolean. Bayesian observation updates; per-node
Shannon entropy; mean-uncertainty + recon-frontier. Black-box = diffuse priors
that sharpen with observation; white-box collapses toward deterministic (MDP).
- pomdp.rs — value_of_information(), decide() (recon vs exploit falls out of
belief entropy), and may_assert() — the mathematical anti-hallucination gate:
no exploitability claim while the belief is diffuse (high entropy) → observe first.
- grounding.rs — verification engine, hard rule "no claim without a tool receipt":
empirical grounding for black-box (raw HTTP/OOB/error markers), symbolic for
white-box (file:line into reviewed source). Ungrounded claims demoted + flagged
receipt_missing (feeds future reward shaping).
- pipeline.finish(): grounding gate before reporting + belief-uncertainty readout.
- bump 3.5.0 → 3.5.1; README documents the v3.5.1 belief/grounding architecture
and the infra/bandit/reward roadmap.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Finding enriched with owasp / mitre / kill-chain stage / exploitability /
business_impact / chains_from (attack-path edges).
- attack_graph module: derive OWASP Top 10 + MITRE ATT&CK technique + kill-chain
stage from CWE (heuristic, no extra model call); render a Mermaid attack-path
flowchart (findings grouped by stage, explicit + implicit edges) and an ASCII
kill chain for the REPL.
- enrich() runs in finish() for every engagement.
- HTML report gains an "Attack Path & Kill Chain" section (Mermaid via CDN, dark)
plus a stage/sev/OWASP/MITRE/exploitability table.
- REPL print_findings shows the ASCII kill-chain + severity summary after a run.
- models: add GPT-5.5, GPT-5.4, GPT-5.4-mini, GPT-5.3-codex, GPT-5.2.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- New GREYBOX mode: review a repo's source AND exploit the running app in one
pipeline — code-review findings become LEADS injected into live exploitation.
CLI: `neurosploit greybox <repo> --url <app> [--creds creds.yaml] [--focus ...]`
REPL: set both /repo and /target → greybox auto-selected.
- Credentials (harness/src/creds.rs, dependency-free YAML subset): jwt / header /
cookie, or an automated `login:` flow. Derives an auth header and/or a
"authenticate first via curl" directive injected into prompts so agents test
authenticated. --creds flag + /creds command + creds.example.yaml.
- RunConfig gains `repo`; run_engagement refactored to a Mode enum (Black/White/Grey).
- Verified offline: greybox loads creds, combines repo+URL, runs pipeline, writes report.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- New persistent interactive session (app/src/repl.rs), launched when run with no args:
banner, model selection, API-key config (/key) or subscription (/sub), then a live
session to set /target, /repo, /auth, and free-text /focus instructions (or just type
them) that STEER which agents run and how.
- Slash-commands: /model /providers /key /sub /target /repo /auth /focus /mcp /votes
/agents /show /run /quit (+ bare text = focus).
- RunConfig gains `instructions` and `auth`:
* instructions bias both LLM agent-selection and the heuristic (focus keywords →
injection/access-control/etc. agents get a strong boost)
* operator directives (focus + auth) injected into recon and exploit prompts so agents
test as an authenticated user and prioritise the requested vuln classes
- bump 3.4.1 → 3.5.0 (CLI, harness, reports, credits)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Task-based model ROUTER (recon/select prefer a fast model; exploit prefers primary; validate uses a different model than the finder)
- ReAct doctrine injected into exploit prompts (Thought→Action→Observation, token-efficient)
- Dedup: unique agents per run + findings deduped by CWE/endpoint/title (highest confidence kept)
- Token economy: recon blob capped for selector + per-agent context
- Configurable MCP: merge user mcp.servers.json into the pipeline's .mcp.json
- +54 white-box/code-analysis agents (NoSQLi, LDAP/XPath, JWT-none, Java/.NET/PHP/Go/Node/Python
specifics, SSTI, ReDoS, deserialization, etc.) → 303 agents total (78 code)
- Credits: Joas A Santos & Red Team Leaders (CLI banner, interactive header, HTML+Typst report)
- README: GitHub stars/forks badges, 60-second quick start, full API config steps, intuitive layout
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
