fix(witness): redact secrets from bundled verify.py output (SECURITY)

The Python proof verifier (archive/v1/data/proof/verify.py) imports the
project settings, which read the user's .env file. When pydantic
validation fails (e.g., extra fields not in the Settings schema), the
error dump includes the offending input_value — which means real
Docker tokens, GitHub PATs, API keys, etc. were being echoed to stdout
and captured into the bundled verification-output.log.

Confirmed on this branch's first bundle generation: dckr_pat_,
tok_... cluster token, and other long opaque strings leaked into
witness-bundle-ADR028-<commit>/proof/verification-output.log inside
the .tar.gz. Bundle + tarball nuked from disk before any push.

Added:
- scripts/redact-secrets.py — stdin->stdout filter with patterns for
  common token prefixes (dckr_pat_, tok_, sk-, ghp_, gho_, github_pat_,
  AKIA, hf_, xoxb-, xoxp-, Bearer), `field=secret` assignments, long
  opaque alphanumeric strings (40+ chars), and long hex runs (20+ chars
  which catch token suffixes after `...` truncation).
- generate-witness-bundle.sh now pipes verify.py stderr through that
  filter before tee-ing into the bundled log.
- Also fixed pre-existing stale `v1/` paths in the witness script
  (correct path is `archive/v1/`).

The user must rotate the leaked credentials regardless (the bundle was
never pushed, but they appeared in this local Claude session log).

Co-Authored-By: claude-flow <ruv@ruv.net>
This commit is contained in:
ruv 2026-05-22 21:04:57 -04:00
parent 4c39e28bd0
commit f8a2e36958
2 changed files with 68 additions and 6 deletions

View File

@ -39,18 +39,18 @@ cp "$REPO_ROOT/docs/adr/ADR-028-esp32-capability-audit.md" "$BUNDLE_DIR/"
# ---------------------------------------------------------------
echo "[2/7] Copying proof system..."
mkdir -p "$BUNDLE_DIR/proof"
cp "$REPO_ROOT/v1/data/proof/verify.py" "$BUNDLE_DIR/proof/"
cp "$REPO_ROOT/v1/data/proof/expected_features.sha256" "$BUNDLE_DIR/proof/"
cp "$REPO_ROOT/v1/data/proof/generate_reference_signal.py" "$BUNDLE_DIR/proof/"
cp "$REPO_ROOT/archive/v1/data/proof/verify.py" "$BUNDLE_DIR/proof/"
cp "$REPO_ROOT/archive/v1/data/proof/expected_features.sha256" "$BUNDLE_DIR/proof/"
cp "$REPO_ROOT/archive/v1/data/proof/generate_reference_signal.py" "$BUNDLE_DIR/proof/"
# Reference signal is large (~10 MB) — include metadata only
python3 -c "
import json, os
with open('$REPO_ROOT/v1/data/proof/sample_csi_data.json') as f:
with open('$REPO_ROOT/archive/v1/data/proof/sample_csi_data.json') as f:
d = json.load(f)
meta = {k: v for k, v in d.items() if k != 'frames'}
meta['frame_count'] = len(d['frames'])
meta['first_frame_keys'] = list(d['frames'][0].keys())
meta['file_size_bytes'] = os.path.getsize('$REPO_ROOT/v1/data/proof/sample_csi_data.json')
meta['file_size_bytes'] = os.path.getsize('$REPO_ROOT/archive/v1/data/proof/sample_csi_data.json')
with open('$BUNDLE_DIR/proof/reference_signal_metadata.json', 'w') as f:
json.dump(meta, f, indent=2)
" 2>/dev/null && echo " Reference signal metadata extracted." || echo " (Python not available — metadata skipped)"
@ -73,7 +73,13 @@ cd "$REPO_ROOT"
# 4. Run Python proof verification
# ---------------------------------------------------------------
echo "[4/7] Running Python proof verification..."
python3 "$REPO_ROOT/v1/data/proof/verify.py" 2>&1 | tee "$BUNDLE_DIR/proof/verification-output.log" | tail -5 || true
# SECURITY: the verify.py emits a Pydantic schema dump on validation failure
# that includes the user's .env contents (Docker tokens, API keys, etc.).
# Redact any line matching common secret-shaped patterns before writing the
# bundled log. See ADR-110 wave 5 incident note.
python3 "$REPO_ROOT/archive/v1/data/proof/verify.py" 2>&1 | \
python3 "$REPO_ROOT/scripts/redact-secrets.py" \
| tee "$BUNDLE_DIR/proof/verification-output.log" | tail -5 || true
# ---------------------------------------------------------------
# 5. Firmware manifest

56
scripts/redact-secrets.py Normal file
View File

@ -0,0 +1,56 @@
#!/usr/bin/env python3
"""Pipe stdin through a secret-redaction filter to stdout.
Used by generate-witness-bundle.sh to strip credentials from log files
before they enter the witness bundle. Pure stdlib so it runs anywhere.
Usage:
some-command 2>&1 | python3 scripts/redact-secrets.py > clean.log
"""
import re
import sys
# Token prefix patterns — common SaaS / VCS API token shapes.
PREFIX_PATTERNS = [
(re.compile(r'(dckr_pat_|tok_|sk-|ghp_|gho_|github_pat_|AKIA|hf_|xoxb-|xoxp-|Bearer\s+)[A-Za-z0-9_\-\.]+',
re.IGNORECASE), r'\1[REDACTED]'),
]
# Long opaque strings (40+ alphanumeric / underscore / dash chars).
LONG_OPAQUE = re.compile(r'[A-Za-z0-9_\-]{40,}')
# Long hex runs (20+ hex chars — covers token suffixes after `...`).
LONG_HEX = re.compile(r'[a-fA-F0-9]{20,}')
# `field=VALUE` style assignment where field name suggests a secret.
SECRET_ASSIGNMENT = re.compile(
r'(token|password|secret|api_key|access_key|private_key|psk|bearer)'
r'(["\'\s:=]+)["\']?([A-Za-z0-9._\-/+]{12,})["\']?',
re.IGNORECASE
)
def redact_line(line: str) -> str:
for pat, repl in PREFIX_PATTERNS:
line = pat.sub(repl, line)
line = SECRET_ASSIGNMENT.sub(lambda m: f'{m.group(1)}={"[REDACTED]"}', line)
line = LONG_OPAQUE.sub('[REDACTED-OPAQUE]', line)
line = LONG_HEX.sub('[REDACTED-HEX]', line)
return line
def main() -> int:
for raw in sys.stdin.buffer:
try:
text = raw.decode('utf-8', errors='replace')
except Exception:
sys.stdout.buffer.write(b'[REDACTED-UNDECODABLE]\n')
continue
sys.stdout.write(redact_line(text))
sys.stdout.flush()
return 0
if __name__ == '__main__':
sys.exit(main())