No description
- JavaScript 77.8%
- Python 10.2%
- Solidity 5.4%
- HTML 5.2%
- Shell 0.6%
- Other 0.5%
| labs | ||
| .gitignore | ||
| AGENTS.md | ||
| CLAUDE.md | ||
| README.md | ||
Unerror Labs
Research repo for security (cyber, physical, social engineering), programming and languages, infrastructure, firmware and IoT, and other related domains.
Each lab is a self-contained folder under labs/<area>/<lab>/ with its own README,
scripts and fixtures. Findings live in the Unerror Labs Obsidian vault
(~/Documents/Vaults/Unerror Labs); this repo holds what is needed to reproduce a lab.
Labs
- Muse agent sandbox — own-instance Muse/Hatch cell-boundary research; probe tools and rerun notes.
- k8s-attack-surface — recent Kubernetes plus the add-ons real clusters install: CVE-2026-2270 reproduced end to end (and the two preconditions the advisory omits), measured stock posture, ingress-nginx at EOL.
- php2shell — CVE-2026-109618/109619 OOM state corruption; exploit-free oracle plus the public reference harness (RCE on stock 8.4.26, fails on 8.4.27).
- Autonomous Bitcoin accumulation — what actually pays sats at $0 capital; bounty radar and mining-margin tools.
- AI opportunity lanes — where the AI risk layer pays: underwriting evidence, agent incident response, and the containment rubric.
- node2shell — CVE-2026-111497 Brotli decoder; reachability/version oracle and a differential fuzzer (no witness found yet).
- oss-risk-radar — which self-hosted AI project to audit next: impact × novelty ranking over exposure, secrets and OSV advisory density (stdlib scorer, public metadata only).
- pydantic-validation-differential — does "validated" mean "understood"? Four-path matrix over pydantic coercion, plus the SSRF negative result that redirected the lab.
- sglang-kv-isolation — is SGLang's
prefix cache a trust boundary under a shared API key? Source-level pass on the
salt model, the
cached_tokensoracle, and the header-override surface. - honeypots — what hits an unadvertised VPS, and what hits it when the decoy looks like AI infrastructure; T-Pot edition choice for 8 GB plus a stdlib decoy for OpenAI and MCP endpoints.
- car-wifi — what the car's hotspot exposes to a passenger: an Android 14 head unit with an open DNS resolver, key-gated ADB on 5555 and paired-only AirPlay, mapped from inside the vehicle.
- data-poisoning — brief: can our own content be made useless to scrapers and detectable if ingested? Tarpit cost measurements, canary provenance, retrieval-time injection.
- smishing — one scam SMS in, a report out: defanged IOCs, link chain, landing-page hash, VT and ASN, shaped for 7726 and the brand/host reports.
- phishing-email — one email in, a report and a campaign key out: SPF/DKIM/DMARC, brand-domain spoof, link-text vs href, attachment hashes with macro flags, and JSONL keys for grouping campaigns.
- scam-bait — engaging a scammer with a persona rather than scanning them: a headless harness that submits fake victim data once and captures the kit's exfil endpoint, plus the free-number problem and a local mock kit.
- decoy-mcp — brief: what an agent discloses when it meets an untrusted MCP tool server — context leakage, cross-server credentials, tool-description injection — and where to place it so only attackers find it.
- seed-sweep — a published BIP39 phrase and four watched addresses: how fast seed-phrase sweepers are, how many are racing, and where the money goes. Unfunded so far, on purpose.
- Agents — meta-lab with a two-fold goal: a useful personal/family agent, and the frontier path (sandboxes, guidance, evals, RL, fine-tuning, model creation) that transfers to work. The build is iris-replacement.
- Family IPTV — Canadian Live TV for Jellyfin via Dispatcharr: curated M3U, EPG grab, dead-stream pruning, plus the Hockey and Matchora Sports Live adapters.
- ETH refutation — OpenAI's claimed subexponential exact-#SAT algorithm: the parameter chain and the one new algebraic gadget, checked locally.
- SSV Network — Immunefi $250k smart-contract bounty: verified which bytecode is actually live (v2.0.0 + May hotfix), the pre-excluded known issues, and where the new effective-balance accounting is thinnest.
- morpho — Cantina's $2.5M Morpho bounty, scoped from the addresses page the rules point at: 795 deployments, 41 audited contract kinds, 1,705 prior findings, and the two gate candidates that died on review.
Start a lab
mkdir -p labs/security/my-lab
$EDITOR labs/security/my-lab/README.md
The README answers: what question the lab asks, how to set it up and rerun it, what was found, and the authorization/scope for anything touching a real system.
Rules
Only test systems you own or have written authorization for. Redact secrets and personal
data before committing. See AGENTS.md for the full conventions.