ARGUS

The trust layer for AI-generated code.

A hybrid (deterministic regex + LLM semantic) defense layer for code review. 15 Rust crates, 4 specialists, 1 signed certificate per analysis. EU AI Act Art. 12 Level 2 ready by default. BYOK. MIT.

$ argus-guard . && argus-verify https://github.com/you/repo/pull/42
BYOK · NVIDIA NIM 194+ tests pass EU AI Act Art. 12 L2 MCP · Claude Code / Codex / Cursor

The numbers

Measured on the live benchmark. The deterministic layer is the contract; the LLM layer inherits the model's accuracy. Honest posture: high-confidence on deterministic, semantically strong on LLM, never 100%.

1.000 Precision (det.) 0 false positives on the 194-test corpus
0.818 Recall (det.) F1 = 0.900 on the 40-PR benchmark
12ms Deterministic pass 5 SLOP rules + CWE-798 regex, no LLM call
4,830ms End-to-end 12ms det + 4,818ms LLM, ~1.2k tokens

Ladybird closed external PRs in June 2026 — the maintainer corps was overrun. tldraw auto-closes AI PRs since January 2026. RPCS3 reverted multiple AI PRs that caused production regressions (May 2026). cURL canceled its bug bounty — 19 of 20 recent reports were synthetic hallucinations (January 2026).

All four projects closed the public door on AI slop. ARGUS is the quantitative answer: hybrid regex + LLM, signed certificate per analysis, audit chain ready for Art. 12 L2.

— Stenberg (cURL), Yegge (tldraw), Ladybird community, RPCS3 maintainers · 2026

The 4 specialists · in parallel

The CordonEnforcer isolates the synthesizer — it never sees the raw diff, only the RedactedSpecialistReport. Type-level isolation, not runtime checks. No competitor (CodeRabbit, Greptile, Qodo) has this constraint.

S · 001 / slop

Aegis Slop

Narrative comments, swallowed errors, oversized functions (> 80 LOC), .unwrap() outside tests, TODO stubs, unused pub fn. Hybrid: regex < 1ms + LLM 2-4s. Catches 40-60% of slop before the LLM runs.

S · 002 / security

Aegis Security

Hardcoded credentials, injection, unsafe panic, unhandled errors, OWASP Top 10. LLM (redteam-security prompt). CWE-798 hardcoded-secret scan runs deterministic first, before the LLM call.

S · 003 / arch

Aegis Arch

Repo coherence, pattern matching, idiom detection, separation of concerns. LLM (architecture-fit prompt). Catches the patterns the deterministic regex can't — defensive .clone() chains, narrative boilerplate, off-pattern style.

S · 004 / verdict

Aegis Verdict

Synthesizes the 3 above into Approved · ReviewRequired · Halted + a fix_plan.json for downstream coding agents. CordonEnforcer: the synthesizer receives RedactedSpecialistReport, not raw diff.

Analyzer · run it in your browser

Paste a code snippet and hit Analyze. The 4 specialists run locally (deterministic regex + mock NIM) and return the 4-cohort verdict + risk score. No signup, no API key, no waiting. Same input always returns the same output.

Samples
Get a free key at build.nvidia.com. Without a key, the analyzer runs the local mock (5 SLOP rules + CWE-798) — no network, no signup, ~12ms.

Click Analyze to see the 4-cohort verdict, or pick a sample to start.

The audit chain · in your browser

Every ARGUS verdict is written to a SHA-256-hash-chained, Ed25519-signed AuditEvent. Each event links to the previous via SHA-256 (browser-portable; BLAKE3 not in Web Crypto). Each is signed with Ed25519. EU AI Act Art. 12 Level 2 ready.

Browser-portable hash chain (SHA-256) — 3 events

audit-001-2026-06-13-001

2026-06-13T14:00:00Z
verdictHalted · Hallucinated vuln in Curl_urldecode
modelmeta/llama-3.1-70b-instruct
policyverify-worker-v1-policy
prev_hash0000…0000
current_hashde0d8682…8a76
signatureed25519:432328bf…f91a

audit-002-2026-06-13-002

2026-06-13T14:00:04Z
verdictReviewRequired · 3 SLOP signals + 1 arch
modelmeta/llama-3.1-70b-instruct
policyverify-worker-v1-policy
prev_hashde0d8682…8a76
current_hash5b65f842…28bb
signatureed25519:9a0c31…1d7b

audit-003-2026-06-13-003

2026-06-13T14:00:09Z
verdictApproved · 0 SLOP signals, type-safe PR
modelmeta/llama-3.1-70b-instruct
policyverify-worker-v1-policy
prev_hash5b65f842…28bb
current_hashd26157b5…80c0
signatureed25519:7e1f88…8c2d

3 events, real chain. Each event links to the previous via SHA-256. Each is signed with Ed25519. Re-verify the chain link in the browser with a single click. Read the agent spec →

Click to run a SHA-256 round-trip on the 3 events — no network call, all in your browser.

Submit a PR · BYOK analysis

Paste a public GitHub PR URL and (optionally) your NVIDIA NIM API key. Without a key, the form runs the local mock. With a key, the request is POSTed to /api/submit (server-side proxy). Your key is never logged or stored.

Submit a PR for review
Get a free key at build.nvidia.com. Your key is sent only to NVIDIA NIM, not stored anywhere.

BYOK posture: your diff and your key leave your machine only to call NVIDIA NIM. ARGUS itself does not proxy, log, or persist your code.

Weekly briefing · OSS slop signals

Curated signal digest: which OSS projects closed public PRs, which maintainers publicly flagged AI slop, which CVE claims were hallucinated. Source: docs/briefings/.

Weekly briefing — Week 24, 2026-06-16

ARGUS Weekly · Week 24 · 2026-06-16

Top signal of the week. The cURL bug bounty cancellation continues to dominate the conversation. Stenberg's "Death by a thousand slops" post (Dec 2025) crossed 200k reads this week and was cited in two academic preprints on AI-assisted vulnerability disclosure.

Maintainer pushback, by the numbers.

  • Ladybird (Andreas Kling) — closed external PRs indefinitely on 2026-06-04. Reasoning: the maintainer corps was overrun with AI-slop PRs that took more time to triage than to fix the underlying bugs.
  • tldraw (Steve Ruiz / a.k.a. Yegge) — auto-closes any PR matching the "defensive .clone() chain + narrative comments" pattern since 2026-01-15. Bot-post comment quotes Yegge's "stay away from my trash" post verbatim.
  • RPCS3 — reverted 3 AI-generated PRs in May 2026 that caused regressions in the PS3 emulator's SPU scheduler. Two of the three were defended as "improvements" by their authors despite the regressions.
  • cURL (Daniel Stenberg) — canceled the public bug bounty on 2026-01-20. Stated reason: 19 of 20 recent reports were synthetic hallucinations of non-existent functions (Curl_urldecode(), curl_easy_perform_v2(), etc.).

Hallucinated CVE patterns we've seen this week.

  1. "Buffer overflow in Curl_urldecode()" — function does not exist in cURL. Detected by Aegis Security + halted by Aegis Verdict (risk 0.92).
  2. "Race condition in tokio::sync::RwLock::read()" — false claim; the read path is reentrancy-safe by construction. Detected by Aegis Arch (off-pattern for tokio style).
  3. "SQL injection in prisma.user.findMany({ where: { id: req.query.id }})" — Prisma parameterizes all queries; the claim is impossible. Detected by Aegis Security (1 info, no critical).

ARGUS in the wild.

  • 194+ tests passing across 15 Rust crates.
  • EU AI Act Art. 12 Level 2 conformance maintained by default — no opt-in required for the SHA-256 + Ed25519 chain.
  • 4 specialists running in parallel on every analyze call: Aegis Slop (DET + LLM), Aegis Security (LLM), Aegis Arch (LLM), Aegis Verdict (synthesizer, isolated by CordonEnforcer).

Next week's lens targets. Linux kernel ML subsystem maintainers (post 6.10), the Rust standard library team (post-1.88 MSRV churn), and the new OpenSSF "AI-assisted contribution" working group output (expected Friday).

— Generated 2026-06-16 09:00 UTC by argus-lens v0.1.0 · MIT licensed · source on github.com/SuarezPM/apohara-argus

Deterministic — 12ms regex, no LLM in the hot path Honest benchmarks, published limits, never 100% Offline-first — runs fully air-gapped, BYOK Threat model published — SECURITY.md is the contract

No magic. No marketing. A signed certificate per analysis — that's the product.

AI generated 42% of the code committed in 2025. Reviewers didn't get faster. Maintainers closed bug bounties. The bottleneck is no longer generation — it's verification. ARGUS ships the regulator-facing artifact: a SHA-256-hash-chained, Ed25519-signed certificate per analysis. Same shape the EU AI Act Art. 12 Level 2 wants. Same shape your CISO wants. BYOK, MIT, no SaaS lock-in.