数据分析

Brainblast Fleet

试用

Autonomous, scheduled agent that finds real, proven security footguns in popular SDKs and submits them to the open Brainblast corpus.

它能做什么

Autonomous, scheduled agent that finds real, proven security footguns in popular SDKs and submits them to the open Brainblast corpus.

技能文档

Brainblast Fleet (OpenClaw)

What this is. Brainblast is an open, verified corpus of real security footguns in popular SDKs — code that compiles and runs but silently does the wrong thing (TLS certificate checks disabled, a zeroed marketplace fee, an unconfirmed blockchain transaction, a skipped signature verification). This skill is the fleet: an autonomous agent that discovers new footguns in public code, proves each one is real, and submits it to the corpus — on its own, on a schedule, unattended.

Each entry is a VTI (Verified Trap Instance): a footgun proven to reproduce RED→GREEN — the checker fires on the vulnerable code and passes on the fixed code. Only proven traps ever land; the deterministic engine decides, and a candidate that doesn't reproduce is discarded, never forced through.

Install it if you want an OpenClaw agent that continuously grows this shared security-training corpus in the background.

What you need

  • git, node (20+), npm, and python3 on PATH. No API keys.
  • ~1 GB of disk for the engine checkout the skill manages at ~/.brainblast/repo.
  • Discovery uses Sourcegraph's public code search — no GitHub token or quota.
  • Optional: BRAINBLAST_INGEST_TOKEN, only if you run your own Brainblast registry and want to lift its 60/hour per-IP submit cap. Not needed otherwise.

Install

openclaw skills search brainblast          # find it on ClawHub
openclaw skills install brainblast-fleet   # (or the scoped @publisher/brainblast-fleet slug shown)
openclaw skills list                       # confirm it loaded

What it does on your machine (and what it doesn't)

  • Clones/updates the Brainblast engine into a managed checkout (~/.brainblast/repo) and resets it to a clean baseline each run. It is sentinel-guarded — it never touches any other repo on your machine.
  • Reads public code only. It searches Sourcegraph and fetches candidate files read-only to prove them locally. It never executes any scanned code.
  • Submits to the public registry at registry.brainblast.tech, which re-proves every submission server-side — so installing and running this contributes findings to the shared, open Brainblast corpus.
  • Holds a single-instance lock, so scheduled runs can never overlap or corrupt the checkout.

When to use

  • The user says "run the fleet", "source VTIs", "grow the corpus", or "scout SDKs".
  • A cron fires the scheduled job this skill installs (see Autonomous cron below).

Quick start (one cycle)

Run the full deterministic cycle with the exec tool:

{baseDir}/scripts/run-fleet.sh

That sweeps every seam. To target specific seams, pass them (--list shows the menu without doing work):

{baseDir}/scripts/run-fleet.sh --list
{baseDir}/scripts/run-fleet.sh objarg absence cstsol

Read the scoreboard it prints: proven-and-promoted vs drafted counts, corpus delta, and any REJECT lines from the registry. Report those numbers back.

Phases (what the script gates, so you can reason about failures)

  1. Reset — the managed clone is reset to its committed baseline (guarded by a sentinel so it never touches a repo it didn't clone), keeping packs/ and candidates/ bounded so prove/SLA never creep past the cron timeout.
  2. Discoversg_scout.py sweeps Sourcegraph seams and writes fleet/candidates/*.json. Each hit carries commit-pinned provenance (repo + SHA + verbatim line), which the registry requires.
  3. Prove + promotenpm run fleet proves every candidate RED→GREEN and auto-promotes the proven into packs/. This is the absolute gate. Candidates that don't reproduce are DRAFT and never advance.
  4. SLA gatenpm run sla must be green (100% of the corpus reproduces) or the script aborts before any submit.
  5. Submit — only the newly-proven candidates are POSTed to the open /api/vti (scoped so the registry's 60/hr per-IP cap is spent on new work, not on re-POSTing the committed corpus). The server re-proves each one and rejects anything without fetchable provenance. Idempotent.
  6. Reprove (optional) — with BRAINBLAST_REPROVE=1 + token, flips new VTIs to proof_verified=true now; the registry also reproves on its own schedule.

If any gate is red, the script stops. Report the failure verbatim — do not try to force it through.

Deepening with sub-agents (optional, for richer footguns)

The seam sweep is broad but shape-bounded. To find novel footguns a seam doesn't cover, fan out isolated sub-agents with sessions_spawn, one per target repo, then sessions_yield to collect them. Give each EXACTLY this contract:

Scout for a **silent integration footgun** in its use of: code where the happy path compiles and runs but does the wrong thing (insecure-default flag, zeroed fee/amount, skipped verification, wrong constant). Shallow-clone read-only (git clone --depth 1). For each real footgun that fits an existing checker (object-arg-property-forbidden-literal is the workhorse), write a candidate to $BRAINBLAST_REPO/fleet/candidates/.json per $BRAINBLAST_REPO/fleet/SCOUT-CONTRACT.md. Capture COMMIT-PINNED provenance: the exact git rev-parse HEAD, provenance.sourceRef = github.com///blob//, provenance.evidence = the verbatim vulnerable line (it MUST contain the trap's propName/call). Self-validate with npm run submit:vti -- --candidate --dry-run --verify-provenance before returning. Do not fabricate — if the repo has no provable footgun, write nothing and report "clean".

Deepening drives the engine directly — not run-fleet.sh, whose managed-clone reset would wipe the sub-agents' candidates. After the sub-agents write to $BRAINBLAST_REPO/fleet/candidates/, from $BRAINBLAST_REPO/packages/core run:

npm run fleet                                   # prove RED→GREEN + promote (the gate)
npm run sla                                     # corpus must be green
npm run submit:vti -- --candidate fleet/candidates/.json   # per proven candidate

Keep sub-agents cheap: set agents.defaults.subagents.model to a smaller model — the proof gate, not the model, decides what lands.

Sub-agents read code; they never execute a scouted repo's scripts. Clone --depth 1, read-only.

Autonomous cron

Install a recurring job that runs a cycle on its own. {baseDir}/scripts/install-cron.sh wraps openclaw cron create with sane defaults (every 6 hours, command payload, no model call):

{baseDir}/scripts/install-cron.sh                 # every 6h
{baseDir}/scripts/install-cron.sh "0 */12 * * *"  # custom cron expression

Inspect and manage it with the standard cron verbs:

openclaw cron list
openclaw cron runs --id 
openclaw cron remove brainblast-fleet

The job runs run-fleet.sh as an isolated command payload — no model call, pure deterministic pipeline — so an unattended run can only ever submit proof-gated VTIs. (Cron sessions can't spawn more cron jobs, by design.) Use the sub-agent deepening path in interactive runs where a model is in the loop.

Report

Summarize: seams swept, candidates proven vs drafted, corpus delta, any rejects, and suggested next seams. Per-run detail is in $BRAINBLAST_REPO/fleet/REPORT.md.

相关技能

按 OWASP Agentic Skills Top 10 审计已安装的 AI Agent 技能,输出文本、JSON、SARIF 或 HTML 报告。

29 次安装1 星标

Create CompleteTech LLC security, safety, permissions, and production-readiness review artifacts for agentic development workflows, including risk intake, to...

11 次安装

Automates agent self-audit by gathering evidence, summarizing runs, proposing minimal patches, and requesting human approval for safe incremental evolution.

11 次安装1 星标

Prompt-injection / jailbreak firewall for AI agents. Scan any untrusted text — a post/comment/DM from another agent, a tool result, scraped web content — BEF...

2 次安装

Habit gate: before EVERY remote skill install, re-fetch, or allowlist promotion, call a hosted heuristic static red-flag scan (shell pipes, base64|bash, quar...

Run your agent against 5 real prompt-injection attacks and get a local report of what it actually did — credential leak, metric fabrication, false-halt, runa...

4 次安装