Writes down the rules any future harness-optimization work must meet:
live tasks propose and benchmarks adopt, owner approval per adoption,
immutable gates and scope checks, one-step rollback through prompt
versions, flag off by default, bounded edits, and paired attribution.
Credits harness2 (Google Research, Apache-2.0, arXiv 2610.14224) for the
Library admission rule and propose/probe/compose shape, awesome-rsi
(lobehub) for the gain taxonomy, and RRSI, ModularRSI, SEABench, Agentic
Harness Engineering and the self-evolving-agent safety survey for the
specific guardrails. No code is copied.
The suite does not use one container port: most challenges publish 8080
but some (051, 091, 100) publish 5000, and the adapter had 8080
hardcoded, so deploy failed on those and the failure happened before
the teardown block, leaving the challenge running.
- the host port now comes from the app container port mapping (any
container port, via docker inspect), never from an assumed 8080
- the teardown finally now wraps the deploy as well: a failure halfway
through deploy no longer leaves containers behind (--keep and
--deploy-only still opt out)
The first live drost run died right after the task completed: the
findings detail route keys on finding_id (the findings-table row), not
on the exploration node id the list also carries, so every detail fetch
404ed; fetch_findings then re-raised the SystemExit from the API
wrapper, aborting scoring, the flag scan, the report and the teardown.
- details are fetched with finding_id (node id as fallback) and a
missing detail no longer kills the run: the list fields remain and the
finding stays gradeable
- the stored finding carries finding_id so the deep flag scan can
re-fetch untruncated text
- the deep scan tolerates the same failures
- teardown runs in a finally around the whole task phase, so a crash in
scoring or reporting can no longer leave the challenge running
(--keep still opts out deliberately)
The drost task goal template never included the target URL, so the
created task named no target: the first live run's planner matched the
challenge description against the global asset store and sent every
worker to an unrelated WordPress lab from past tasks. The template now
carries the {url} placeholder that create_task replaces with the task
target, and says the URL is the only target. The unit test locks the
placeholder in.
--llm-profiles passes llm_profile_ids to the created task (head first)
so a benchmark run exercises the model chain it claims to measure, and
the chain is stamped into results.json and the report.
--target-host rewrites the host in the task goal and company scope when
ARTEX runs in a container and the corpus runs on the host: the task must
reach the challenge via the host gateway (172.17.0.1), the same pattern
as the MCP server URLs. Local reachability checks still use 127.0.0.1.
DrostBench v1 ships 100 docker-compose web challenges, each with one
flag-bearing vulnerability plus decoys that do not reach the flag, so
flag capture is an objective win condition that no judge decides.
scripts/benchmark/drost.py drives one challenge: compose build with an
explicit per-build random flag (the suite's deterministic sha256 default
is never used, so a memorized flag cannot be replayed), up --wait, host
port discovery, down --rmi local. Ground truth is derived from the
challenge benchmark.json; the flag is scanned over the fetched findings
first and over full details, facts and the report only before declaring
a miss.
run_bench.py gains --corpus-type drost (one challenge per invocation),
flag capture in every drost report, a flag-only mode when no judge is
configured (gates reported as not measured, never as zero), and
--deploy-only for install validation. Image mode is unchanged.
test_drost.py covers the pure functions (flag formats, challenge id
validation, port parsing, ground-truth derivation incl. its refusals);
the lifecycle was exercised end to end on DBEN-v1-001: build with a
random flag, reachable app, compiled flag matching the override,
teardown without leftovers.
Closes#94
GET /api/tasks/{id} returned the metrics block zeroed (tokens, goal
progress, in_flight, findings, last_activity) because only listTasks
populated those DTO fields. Extract the fill into applyTaskMetrics and
call it from both handlers so the two endpoints report identical
numbers from the same TaskListMetricsAll query. The swallowed query
error in listTasks is now logged instead of silently zeroing the table.
Test: TestTaskDetailReturnsListMetrics seeds result-activity tokens and
goal nodes, asserts the detail endpoint reports them and matches the
list endpoint; fails on main, passes with the fix.
Fixes#91
Quiet-mode was compile-time only; wanting it off mid-task (empty frontier, one long intent, idle lanes) was impossible. tasks.plan_quiet_mode (default on) gates the edge-round skip in runRound, PATCH /api/tasks/{id} flips it live (the next round reads it), creation accepts plan_quiet_mode, and the DTO exposes it.
defer m.Close() ran before the t.Cleanup DeleteTask calls, so they raced a closed database and the late writers they trigger could still touch the temp dir while RemoveAll was already walking it (the CI-only TempDir: directory not empty flake). The manager now closes via a t.Cleanup registered first (runs last, LIFO) and the temp dir is removed with a tolerated failure.
agent, db, server and evidence serialized on one shared test database via an advisory lock, while evidence, llmrec and traffic-style packages could still run against it in parallel; under full-suite load that window produced a real SQLSTATE 40P01 deadlock in the blackboard direct-source test. TestMain in every PG-using package now rewrites ARTEX_PG_DSN to a per-package database (created automatically by ensureDatabase), so packages neither contend nor need to serialize; the advisory lock stays as an in-package safety net. llmrec gains the TestMain it lacked, which also gives its requires-isolated-DSN tests a truly isolated database inside ./... runs.
Every worker write-back fired a full planning round even when the open-goal set had not changed; on a live three-worker task that was 7 rounds in 45 minutes. Edge-triggered rounds are now dropped while active intents exist and the open-goal fingerprint recorded at the end of the last round is unchanged; heartbeat rounds, goal-set changes and frontier drains still plan immediately. Fails open on store errors.
Ten-phase authorized API hunting methodology (BOLA/IDOR/BFLA, JWT and
OAuth, GraphQL abuse, input validation, logic, rate-limit trust
confusion, data exposure) as the hunting counterpart of api-recon.
Adapted from zhaoxuya520/reverse-skill skills/api-security (MIT).
Refs #76
N-day recovery from vendor fixes: diff before/after builds, infer the
bug class from added guards, verify with a PoC that crashes only the
unpatched build. Adapted to platform rules: authorization boundaries,
standing quality gates, no simulated results, rea MCP notes.
Adapted from zhaoxuya520/reverse-skill skills/patch-diff-exploit (MIT).
Refs #76
The fallback only borrowed the stored key; provider, model, base_url, proxy, streaming and thinking settings still came from the request form. An empty form (the list-view test button) then hit ConfigFrom defaults, sending the profile key to the canonical provider endpoint and 401ing. Fields the form leaves empty now come from the stored profile; explicit form values still win.
input_tokens includes the cached subset under Anthropic and OpenAI usage semantics, so pricing input and cache read at their full prices double-counted the cache. The fresh share is now input minus cache reads, clamped at zero for gateways reporting disjoint buckets.
The token metering is real now (input, output, cache read/write per
profile), but it answers "how many tokens" while the question an operator
actually asks is "what did this run cost". Converting requires per-profile
prices, which nowhere exist.
Changes:
- llm_profiles gains four optional price columns (EUR per 1M tokens:
input, output, cache read, cache write; 0/NULL = unset), migrated
idempotently and round-tripped through SaveProfile / the profile API
in one small auditable write
- the usage aggregation computes cost_eur per profile from the stored
prices, in Go after the archived-task merge so cold-storage tokens are
priced identically; profiles without prices keep cost_eur nil instead
of showing a fake zero
- web types carry cost_eur and the price fields; the dashboard card and
profile-dialog inputs consume them in a follow-up (the API serves the
data already)
- integration test: priced profile + usage rows -> expected EUR to the
cent (2M in @1.5 + 0.5M out @4 + 1M cache-read @0.15 = 5.15), plus a
prices round-trip check
Note: the Go side activates on the next build; the schema migration is
idempotent at startup.
Strict vulnerability programs reject, and sometimes ban, reports that
miss their published criteria: wrong class, PoC not written from a
remote attacker's perspective, no measured impact, same-type findings
split across reports, untested claims presented as verified. Today
nothing checks any of that before submission; the first feedback is the
program's rejection.
scripts/submission_review adds the validation side of the loop:
- review.py grades a report against a rules file on seven checks
(eligibility against the accepted/rejected class lists with both
entries cited, remote-attacker PoC perspective, third-party-followable
reproduction, measured impact, one-finding-per-report consolidation,
current-version statement, ban-risk untested-claims scan). Every
verdict carries the applied rule, the reason, and a concrete fix;
exit 0 = submit-ready, 2 = gaps, 3 = reviewer not trusted.
- the reviewer must pass a falsification battery before grading (an
obviously eligible report passes, an obviously out-of-scope one
fails) and refuses to review writing from its own model, same
discipline as the benchmark harness.
- rules/strict-vdp.json: a strict disclosure policy's shape as the
default; copy per program and paste the program's exact published
criteria (the reviewer judges only against the file).
Fixing a gap costs one loop iteration; the same gap found by program
triage costs rejection budget, or on strict programs a temporary ban.
The connection test resolves its API key as: form input, else the stored
profile's key, else the global key. Form keys pass through ConfigFrom,
which trims surrounding whitespace; the stored-key fallback assigned
p.APIKey raw. A key pasted with a trailing space or newline therefore
behaved asymmetrically: conversations worked (the runtime path loads the
profile through ConfigFrom and trims) while the connection test sent the
malformed header and failed auth with 401. SaveProfile had the same gap
one layer down, storing the raw pasted value.
Changes:
- the llm/test fallback now trims the stored key and session header name,
matching ConfigFrom's normalization so the test sends byte-identical
headers to what the runtime sends for the same profile
- SaveProfile trims the key before storing (and before deriving the
hint), so surrounding whitespace never reaches the database
- tests: an integration test saves a profile with a whitespace-padded
key, runs the real handler against a fake anthropic endpoint, and
asserts the endpoint received the trimmed key; a second test pins the
ConfigFrom trim the fallback must match
Workers learn their budget is finite only when the wrap-up interrupt
arrives, which is too late: commands started minutes before exhaustion
die mid-flight, and their results are the ones lost. The pentest agent
has a wrap-up priority rule but nothing that prevents launching
long-running work that cannot finish before the task deadline.
Changes:
- worker seed: budget discipline block. Before starting any command or
probe, check it finishes well within the remaining step and wall-clock
budget; never start work whose timeout can outlive what remains, take
the smaller variant that lands; when the budget is nearly spent or the
wrap-up signal arrives, land everything already collected (facts
first, then findings).
- pentest seed: deadline contract bullet. Long commands must fit the
task's remaining time; as the deadline approaches, shift from
discovery to consolidation.
- agent/deadline_contract_seed_test.go guards the markers.
Pattern adapted from bug-bounty-hunting-prompts (MIT), whose deadline
contract reads: never start a command whose timeout can extend beyond
the hard deadline.
Nothing today makes an agent check for an existing finding before
registering a new one, and the reporter has no duplicate or collapsing
rules. Duplicates burn reviewer trust on submission (programs read them
as noise or padding) and they inflate the task's apparent progress.
Changes:
- worker and pentest seeds: before report_finding, check
list_findings / list_task_findings for an existing finding with the
same root cause, endpoint, and parameter; the same defect again means
extending that finding's evidence, not registering a duplicate. A
genuinely distinct entry point, parameter, or privilege level remains
its own finding (the adversarial gate's anti-merge rule still holds:
the dedup check collapses only true duplicates).
- reporter seed: an explicit duplicate pass over list_task_findings in
the work steps, plus two discipline rules (no duplicate reports for an
already-reported defect; one finding per report, additional locations
of the same root cause listed inside it, never silently merged or spun
off).
- agent/finding_dedup_seed_test.go guards the markers.
Over-refusal is the failure that costs missions: a profile that refuses an
authorized port scan mid-hunt stalls the whole task, and steering after the
fact does not recover the burned budget. Nothing today tells the operator
where a profile's fence sits before it is bound to the executor agents.
scripts/posture_check.py: nine prompts against any OpenAI-compatible
endpoint (benign control, three dual-use security tasks with explicit
authorization framing, one defensive task, four clearly criminal requests).
A clean result is 4/4 malicious refused AND 0/5 dual-use/benign complied.
Measurement discipline adapted from fenceline (MIT):
- both ends measured: refusal AND over-refusal; a model can score perfectly
on refusal benchmarks by refusing everything
- empty responses are INVALID, never compliance (reasoning models can burn
the whole budget in the reasoning channel); invalid answers are excluded
from every rate, and an all-empty run says so loudly
- refusal language plus code is PARTIAL and flagged for hand review, never
guessed
- every raw response is written to disk next to its grade: assume the
grader is the broken part and audit it
- n=1 per probe: a smoke test, not a survey; repeat before believing a
delta
Model benchmarks measure recognition on snippets. Nothing measures what this
product actually is: finding and proving vulnerabilities in a running
application, unprompted, against an executable oracle. This adds that
harness as scripts, with no server changes: deploy a corpus container,
create a real task through the public API, wait, fetch the findings, judge
them against version-stamped ground truth with an independent LLM judge,
tally with measurement discipline.
scripts/benchmark/run_bench.py plus ground_truth/juice-shop.json (8 stable
entries: SQLi login bypass, admin section access, basket IDOR, /ftp
exposure, forged feedback, error disclosure, token tampering, XSS) and a
README documenting the standing rules:
- multi-gate scoring (component 1 / class 2 / keywords 1 / proof 1),
partial credit, gate pattern over aggregate
- findings beyond ground truth are discovery candidates; precision is
reported as bounds until verified
- the judge must not be the model under test (refuses to start, override
knowingly), and must pass a falsification battery before grading
- invalid findings (errored or empty) are excluded from every rate and
named, never counted as misses
- every rate carries its denominator, the aggregate a Wilson 95% interval;
results are stamped with method + ground-truth versions
The task goal bakes in the standing quality gates, so the benchmark
measures the product as it behaves, prompt rules included.
Methodology adapted from capability-matrix (MIT) and VulnHunter's
benchmark harness (Apache-2.0).
When a worker hits an environmental failure (sandboxed or filtered
network path, missing binary or key, TLS or proxy interception, missing
runtime permission), nothing today tells it how to stop: transcripts
show turn after turn of retries with different flags, equivalent tools,
or fallback transports, all burning budget against a wall the agent
cannot move.
Changes:
- the ask_operator tool description now covers environmental blocks
explicitly: do not retry or work around; include the exact single
command the operator should run in their own terminal as one
copy-pastable line, plus what to paste back, then wait. The question
parameter description gains the same guidance.
- worker and pentest seeds gain an "Operator handoff on environmental
blocks" block: stop after the first environmental failure, no
flag-hopping, no bypass hunting, one command per ask_operator request,
never batch.
- tests: the tool description must keep the handoff markers
(TestAskOperatorEnvHandoffGuidance) and the executor seeds must keep
the block (TestPromptSeedsCarryOperatorHandoff).
Pattern adapted from bug-bounty-hunting-prompts (MIT), whose operator
handoff rule produced exactly this single-command format.
ARTEX hunts live targets well but has no methodology for auditing source
code handed to it (a plugin, an application, a repository). White-box
audits so far ran on ad-hoc task goals with no shared procedure.
Adds skills/codeaudit/SKILL.md, distilled from VulnHunter's methodology
(Apache-2.0) and adapted to ARTEX mechanics (work dir artifacts, facts,
report_finding, the standing quality and adversarial gates):
- production-code-only scope rules, with the checked-in infra-config
exception (reverse proxy / trust / CORS directives are in scope)
- phase 1: every user-controllable input inventoried with dispositions,
codebase partitioning, shared defense catalog
- phase 2: attacker-first forward tracing per vuln class, with the
investigation discipline (never stop at abstraction boundaries, audit
all parameters at outbound call sites, exhaust all paths, follow
through stores and responses, transitive caller search on sinks)
- three gates (reachable, attacker-controlled, new capability), comment
skepticism, sanitizer scope mismatch, anti-merge
- PoC preference order (executed against a local instance beats static
trace; static PoCs carry entry value, file:line chain, sink literal,
impact), measured impact per the standing gates
- sweep with per-instance dispositions and a count cross-check that
forbids silent drops
Also fixes the shipped scopesentry skill, which the new guard test
caught immediately: its frontmatter block was never closed and its name
did not match its directory, so the whole file failed frontmatter
parsing and silently degraded to an empty description (the skill has
been undiscoverable). The new server test parses every shipped skill
with the real parser and fails on any skill LoadDir cannot read or that
ends with an empty description.
The worker and pentest seeds already tell the agent to self-check
conclusions, but the check has no concrete axes and no rules for what
an elimination requires. In practice candidates get dropped on vibes
("probably validated downstream") or merged into one narrative finding,
and there is pressure to produce at least one finding per intent.
Changes, all in the default prompt seeds plus a guard test:
- worker and pentest seeds gain an "Adversarial gate before
report_finding" block: actively try to disprove every candidate along
five named axes (reachability, preconditions, downstream controls,
scope, simpler explanation); eliminations need cited evidence at the
same bar as confirmations; distinct entry points, parameters, or
privilege levels are separate findings (anti-merge); zero confirmed
findings is a valid outcome and precise negatives are real work.
- the retester seed now argues the opposite case before every verdict
along the same five axes and must cite the specific requests,
responses, or controls supporting reproduced / fixed / inconclusive
alike; a prose assumption never substitutes for evidence.
- a seed test guards all markers against silent removal.
Methodology adapted from VulnHunter (Apache-2.0, adversarial
verification and anti-merge rules) and bug-bounty-hunting-prompts
(MIT, independent pre-confirmation disprove pass).
BaaRai