Agency OS Radar

Daily open-source discoveries for a one-person AI software agency

Report date: Thursday, November 5, 2026 · Agency OS Edition · Addis Ababa (EAT)

Start Here (plain English)

The problem: vibe coders ship fast — and speed is exactly how real API keys, passwords and tokens end up in code, commit history, or client deployments. Worse, "I'm pretty sure nothing leaked" is not proof. A one-person agency that can't show security evidence loses clients' trust the moment something goes wrong.

The solution in today's radar: one tool, one job, done properly. Betterleaks — the new secrets scanner from the original Gitleaks author — doesn't just find leaked credentials; it can validate whether they are actually live and analyze what identity and permissions they grant. That turns "trust me, it's safe" into a checkable artifact your gate engine can consume before release.

One featured repo per day, focused on a main vibe-coder struggle. All links and the license were verified from the repo's own files on Nov 5, 2026; star counts are point-in-time snapshots from public trackers, not invented; anything unverifiable is marked UNVERIFIED. The business-opportunity section is retired from this format.

The Featured Tool: Betterleaks TEST NOW

Link: github.com/leaktk/betterleaks (the README references the canonical org, betterleaks/betterleaks)
License (verified): MIT — verified from the LICENSE file today.
What it does (simple English): a fast, configurable secrets scanner — API keys, passwords, tokens — for filesystems, git repos, URLs and GitHub orgs. Maintained by the folks who made Gitleaks, including the original author.
What changed recently (verified): v2.0.0-rc.5 released Oct 28, 2026 (v1.9.0 was the Sep 29 stable); v2 stable looks close. Trending in security discussions Nov 3–4. Ships in MegaLinter; used by Aikido Security and others (per README). Roughly 2.7k stars (tracker snapshot, Nov 3).
Why it matters to a vibe coder: Betterleaks adds what Gitleaks lacked: validation (is the secret actually live?) and analysis (what identity and permissions does it grant?) — plus optional revocation and Expr-based filtering to cut false positives on generic strings like password = "...". It turns "we found something" into "here is the blast radius."
Costs: MIT, free, no subscription. Detection-only by default — no outbound requests unless you pass -v/-a. Open source ≠ free model access, but this tool needs no model at all.
Windows compatibility & installation (verified paths): go install github.com/betterleaks/betterleaks/v2@latest (needs Go) or Docker (ghcr.io/betterleaks/betterleaks:latest); brew/dnf cover macOS/Linux. Prebuilt Windows binaries UNVERIFIED today. No special hardware.

The Task Map: proving no secrets leaked, mapped to Agency OS

Lifecycle stage	What Betterleaks does here	Artifact produced	Which gate consumes it
SECURITY (secrets domain)	Scan source + git history for credentials; optionally validate liveness and analyze blast radius	Findings with rule IDs, validation status, severity + exit code	gate engine SECURITY → RELEASE transition (additive)
RELEASE	Blocking pre-push check: no release with unreviewed findings	prepush_release_check.ps1 exit code + findings report	RELEASE gate
MAINTENANCE	Recurring scans of the repo and its full git history on client projects	Scheduled scan reports; diff against last clean scan	Maintenance evidence log
LEAD → DEVELOPMENT	Not covered by this tool (honest gap)	—	—

A focused tool, not a factory: it strengthens the SECURITY → RELEASE → MAINTENANCE slice of the lifecycle and leaves every other stage untouched. It complements gate_engine.py; it never replaces it.

Honest Gaps

Not covered: dependency auditing, code vulnerabilities, authn/authz review, config review — Betterleaks is secrets-only. Your existing security checks (pip-audit, npm audit, Trivy) remain needed for those domains.
Agency OS compatibility: plausible, untested with your scripts — marked as such until you run the experiment below.
v2 is a release candidate: pin the exact version in your pipeline so a moving RC can't change gate behavior overnight.

Why It Fits Agency OS

No invented facts: detection-only by default; validation status is a first-class field, so "live" vs "detected-only" is explicit, matching your REAL/SIMULATED/VERIFIED/UNVERIFIED discipline.
Evidence before advancing: the scanner's exit code and findings list is exactly the machine-checkable artifact a gate needs; a "clean" claim without it stays UNVERIFIED.
Human approval points: waivers of findings and any validation/revocation runs require explicit human approval — the tool keeps dangerous actions opt-in.
Local-first: runs on your Windows PC, no cloud dependency, no data leaving your machine during detection scans.

Limitations, Risks & Human Review

Validation (-v) makes real HTTP requests to credential providers — only run it against credentials you own, with authorization.
Revocation can destroy credentials — never run it without explicit human approval.
Findings describe weaknesses: keep reports private; false positives are possible — verify before reporting to a client.
Scanning a repo's full history needs read access to that history — only scan code you own or are authorized to scan.

Test Today: Betterleaks on Windows (~20 minutes, no subscription)

Prerequisites: Windows PC with Go (go.dev) or Docker Desktop; ~500 MB free space. Internet only for install. No API keys, no model access.
Official installation steps (from the README): go install github.com/betterleaks/betterleaks/v2@latest — or docker pull ghcr.io/betterleaks/betterleaks:latest.
Practical test: in a scratch git repo you own, plant a well-known fake key (e.g. AKIAIOSFODNN7EXAMPLE — the AWS docs example, never a real one) plus a realistic DB_PASSWORD=.... Then run betterleaks fs . and betterleaks git . (full history). Do not add -v/-a in the first run — detection-only is the safe default.
Expected results: both planted secrets flagged with rule IDs and severity; a clean exit code on a clean folder. That output is your SECURITY-stage evidence.
Safety precautions: scan only code you own; never validate credentials you don't control; never run revocation; keep reports private.

Agency OS Integration Proposal: Betterleaks as SECURITY-gate evidence

Problem solved: the SECURITY stage has no machine-checkable proof that secrets were scanned before RELEASE.
Workflow/agent improved: security agent's scan step, feeding — never replacing — gate_engine.py and prepush_release_check.ps1.
Lifecycle stage: SECURITY → RELEASE.
Inputs: project folder + git history. Outputs: findings with rule IDs + exit code.
How results are validated: the gate script counts findings and records the run as evidence in project-state; a verdict of "clean" without a scan artifact stays UNVERIFIED by your own rules.
What happens on failure: findings present → marked BLOCKED, report attached, no transition to RELEASE.
Gate or human approval? blocks the gate automatically on findings; any waiver requires human approval; validation/revocation steps require explicit human authorization.
Smallest safe experiment: in a scratch project, add one betterleaks invocation to prepush_release_check.ps1, plant the fake AWS key, and confirm the check blocks. Additive only — no existing gate logic changes.

Daily Learning: detection vs validation of secrets

Concept (simple English): a scanner detects that a string looks like a key (pattern/entropy). Validation checks with the provider whether it still works. Detection is safe and offline; validation is a live network request that reveals what the credential can do.
How an AI agent might get it wrong: it declares "this leaked key is critical!" (no validation — maybe it's revoked) or "that finding is a false positive" (no evidence), or worse, runs validation/revocation on credentials it doesn't own.
How you recognize the failure: a severity claim with no validation status or scan artifact behind it; unauthorized outbound validation attempts in logs.
What detects it: Betterleaks' detection-only default plus opt-in validation — the artifact shows which stage produced each claim. Agency OS rule: security claims are UNVERIFIED until a scan artifact exists.

Also on the Radar This Week (verified links, brief)

Octomind (Apache-2.0, trending Nov 3–4) — Rust coding-agent runtime with TOML guardrails; policy-as-code inside the agent harness. Windows installer is a remote script — review before running. INVESTIGATE.
Worktrunk (MIT/Apache-2.0, trending Nov 3–5) — git worktrees for parallel AI agents; Winget installs it as git-wt. TEST NOW.
mattpocock/skills (MIT, trending Nov 3–5) — small, composable engineering skills; no process takeover. TEST NOW (read first).
NVIDIA Auto Ontology (Apache-2.0, trending Nov 3–4) — semantic layer so agents query the right data; needs NVIDIA NIM key + Docker. INVESTIGATE.

Final Summary

Best discovery today: Betterleaks (MIT, v2.0.0-rc.5 on Oct 28) — secrets scanning with built-in liveness validation, from the original Gitleaks author. Verdict: TEST NOW.
Best security/reliability improvement: Betterleaks' validate/analyze pipeline — proof, not guesses, about leaked credentials.
Best Agency OS integration: Betterleaks findings + exit code as SECURITY-stage evidence blocking RELEASE via the existing gate engine — additive only, waivers stay human-approved.
One practical action today: install Betterleaks, run a detection-only scan over your Agency OS repo's git history, then wire the check into prepush_release_check.ps1.