AI-generated Django code compiles and handles the happy path. It also tends to skip Django's own security layer:.raw with f-strings, missing CSRF_COOKIE_SECURE, DEBUG=True in prod, mark_safe on user content. BrassCoders' 12-scanner stack catches all four before deploy.
Every claim BrassCoders makes has a citation. Now the whole research index is public on GitHub: 45 sources, 8 categories, covering AI-code CVE rates, package hallucination, credential leakage, and dependency confusion. Machine-readable JSON included. CC BY 4.0.
Cursor writes code fast. Reviewing 300-2,000 static-analysis findings by hand isn't feasible for anyone.
Run brasscoders scan., open.brass/ai_instructions.yaml in Cursor, ask it to triage. Deterministic scan first, reasoning second. Under 5 minutes.
Delete the function under test. If your test still passes, it never tested anything. AI assistants write these constantly: assert response is not None passes on a 404 too.
BrassCoders flags always-true assertion patterns via Pylint.
A local brasscoders scan is voluntary. A CI gate isn't. Drop brasscoders scan. into GitLab CI, upload.brass/ as an artifact, and enable "Pipelines must succeed" so a CRITICAL finding blocks the merge button, not just your conscience.
AI models keep getting more fluent. Clean variable names, idiomatic structure, senior-engineer style. But Veracode's 2026 report found 45% of AI-generated code still ships an OWASP Top 10 flaw on first generation. Clean code just makes reviewers wave the bad line through faster.
12 scanners, free, find every bug: 1500+ raw findings on a typical codebase. That was never the problem. Nobody reads 1500 findings. BrassCoders Paid ranks that pile down to the 50-80 that matter, for $12/dev/month. Same bugs. Shorter list.
Your AI coder writes as much JS/TS as Python, and the same bugs ride along: hardcoded tokens, unsafe patterns, phantom npm packages. Brass runs a Babel-based JS/TS scanner automatically, same scan as Python, findings tagged in one.brass/ output.
The normal way to scan Python: install Bandit, Pylint, and Semgrep separately, wire up three configs, reconcile three output formats. BrassCoders runs six engines (those three plus Pyre/Pysa, ast-grep, detect-secrets) in one pass, merged into one ranked YAML.
Same 12 scanners, free or paid. Both find the same bugs. The OSS core dedupes ~1500 raw findings down to ~300, free forever, zero network calls. Paid ranks that to 50-80 for $12/dev/month. Pay when triage time costs more than that. Not before.
AI assistants write test fixtures with realistic names, emails, phone numbers, sometimes SSNs. Some are invented. Some are real data memorized from training and reproduced verbatim. Either way it gets committed and lives in git history.
AI assistants write O(N2) loops that pass every test and crawl at scale: quadratic string concat, list.insert(0) in a loop, triple-nested joins, unbounded while-True reads. In our benchmark, Bandit, Semgrep, and Pylint caught zero.
AI assistants hardcode realistic-looking API keys into example code, and you copy the key right along with it. brasscoders bundles detect-secrets with a custom layer covering 20+ credential types, redacting every match before it's written anywhere.
A 2025 USENIX study found 19.7% of AI-recommended packages don't exist. Lasso Security registered one hallucinated name as a test and it pulled 30,000+ downloads. BrassCoders checks every import against the registry and flags what doesn't resolve, before pip install runs.
An LLM reviewer can read the same file twice and return two different verdicts. A deterministic scanner returns the same findings every commit. A 2025 benchmark found LLMs win on recall but mislocate the line, static tools nail it. Run both, deterministic first.
Scanned 15 AI-generated Python files. Every single one got a finding, 53 total, 16 critical or high. After triage: 9 of 15 (60%) had a real security or performance bug. The rest were noise, like loop counters flagged as O(N^2) string concat.
The gap is the point.
A 2025 survey sorts AI-code bugs into eight classes. Only some have a structural marker a scanner can match on.
brasscoders catches hallucinated imports, hardcoded secrets, perf anti-patterns, and syntax errors deterministically.
METR's 2025 trial: experienced devs were 19% slower using AI tools, but felt 20% faster. That gap is where unreviewed bugs live. Brasscoders runs the same deterministic scan on every commit, no matter how confident anyone feels about the code.
The 2025-26 data on AI code: 45% of AI-generated code ships an OWASP Top 10 flaw (Veracode). AI-assisted PRs carry 1.7x more issues (CodeRabbit). ~20% of AI-suggested packages don't exist (USENIX).
Brasscoders runs 12 deterministic scanners to catch this pre-merge.
One scan writes.brass/ai_instructions.yaml. Plain text, no proprietary format. Claude Code reads it automatically. Cursor and Continue pick it up with @file. Same severity-sorted findings, line numbers, and fix notes in all three. No plugin needed.