Defensive recon on assets you own. Give it a target endpoint you operate and the
repositories to check, and the swarm proves which of this week's published CVEs
are actually reachable in your stack — reading code, never exploiting it — then
opens a dependency-only pull request for the ones that are. It is a deployable
OpenComputer serverless-agent template, launched from the ?security=1 flow in
the app or run directly from a session.
This is a scanner for your own code and endpoints. It is passive and static by construction (see Scope and safety); it is not an exploitation tool. You confirm you own the target with a declaration at launch, and the swarm can only touch repositories your GitHub App install grants.
scan() passive fingerprint of a host you own (headers, banners, public files)
│
prove() the controller clones each repository and proves reachability per CVE
│ by reading code — never by exploiting it
│
fix() the smallest version-pin bump for each EXPOSED finding → guarded PR
│
report() one structured findings set to the bound destination
The verdict for every repository × CVE is one of:
- exposed — the affected version is pinned and the affected surface is reachable. The controller names the single version pin that fixes it.
- not-exposed — the product is absent, out of the affected range, or provably never built/served.
- needs-review — evidence is ambiguous. Reported honestly, never rounded to "safe".
The controller sweeps the (repository × CVE) matrix sequentially inside one
session, using the sandbox shell to clone and read each repository. It is one
model reasoning over the whole matrix, not N parallel worker processes — that
parallel fan-out is a planned fast-follow. The pipeline, the passivity, and the
dependency-only fix guard are identical either way.
SEED_CVES in feed.ts mirrors the
queue on the launch page: PHP2Shell (CVE-2026-109618), GitLab CE/EE
(CVE-2026-110264), Windows Shell (CVE-2026-96951), and the Oracle 10.0s
(Access Manager CVE-2026-21445, SOA Suite CVE-2026-21446, E-Business Suite
CVE-2026-21447). The launch payload usually names the exact CVEs to run; every
entry is validated by parseAdvisory before the swarm acts on it.
There is no hardcoded allowlist to maintain. Scope is two things:
- The GitHub App installation. The swarm can only clone or open pull
requests against the repositories that install grants. Select fewer there to
narrow scope.
gitandghuse the injectedGH_TOKENinside the sandbox — no token to paste, none in a URL. - Your declaration. At launch you attest you own the target and are
authorized to test it. The scan stays passive in code regardless (see
config.tsandscan-surface.ts).
Keep PUBLISH_ENABLED false (the default) for a first run — fix() still
verifies and reports the exact dependency-only branch and diff, it just writes
nothing to GitHub.
npm install
npm test
npx opencomputer template validate .
opencomputer link --create-project cve-swarm
opencomputer github connect # install the GitHub App on your repos
opencomputer deploy- Model: set the
SWARM_MODELruntime variable (defaults toGLM-5.3). - Findings destination: bind the
securityoutbox to a Slack conversation (or other channel) in the OpenComputer dashboard. The agent publishes one structured findings set per run; it never picks the destination itself. - Schedule:
watch-advisories.tsruns daily at 13:00 UTC. Adjust the cron to match how often your feed updates.
Launch a swarm by sending the controller a payload: the target endpoint, the
repositories, the CVEs, and your ownership declaration. The ?security=1 flow in
the app composes this for you.
Narrowness is enforced in code, not just documented:
- Passive scanning.
scan_surfaceissues plain GET requests for a small, code-owned set of public paths and reads banner headers. It sends no payload, follows no cross-host redirect, speaks https only, and never probes authenticated surface. Seescan-surface.ts. - Static reachability. The controller proves reachability by reading code and configuration with the shell. It is instructed never to run the project, start a server, or trigger a vulnerability.
- Guarded fixes.
open_fix_pull_requestclones fresh and applies exactly the one version pin named — one file, onefrom→tosubstitution, one changed line — or refuses. Publishing is gated byPUBLISH_ENABLEDand defaults to a dry run. Seeopen-fix-pr.ts. - Untrusted evidence. Advisory text, response headers, and repository files are treated as data, never as instructions, and can never widen scope or change the outcome rules.
opencomputer/
agents/cve-swarm/
agent.ts controller: scan → prove → fix → report (model-driven)
config.ts code-owned safety rails (passivity, bounds, publish gate)
feed.ts this week's CVE seed list + validation
authorization.ts input normalization (host/repo), pure and tested
tools/ scan_surface, open_fix_pull_request, report_findings, github
outboxes/ schedules/
channels/ outboxes/ findings delivery
test/ feed, authorization
npm install # installs the agent deps too (postinstall)
npm test # unit tests for the pure logic
npm run build # typecheck