Expected classes of findings

• SQL injection via string concatenation into a query.
• Command injection via os.system/subprocess shell=True with user input.
• Path traversal on a file-read endpoint.
• Hard-coded secret or SECRET_KEY.
• Possibly: insecure deserialization, open redirect, or reflected XSS.

Expected findings

• JWT verification with algorithm: 'none' accepted or signature skipped.
• IDOR: user-controlled ID used directly in a database lookup with no ownership check.
• NoSQL injection via object in request body passed straight to Mongo.
• Mass assignment: Object.assign(user, req.body) enables privilege escalation.

Expected findings

• Stack buffer overflow via strcpy/gets/sprintf on fixed-size buffer.
• Timing attack on strcmp or memcmp password check.
• Weak RNG (rand()) used for token generation.
• Integer truncation in a size calculation.

At least one finding should be exploitability-rated "likely RCE" given sufficient size controls.

Expected findings

• Database password or API key in plaintext.
• debug: true or equivalent flag left on for production.
• CORS wildcard or TLS verification disabled.
• Default admin account or weak default credential.
• Internal hostname or private IP referenced directly.

Expected cross-file chain

A secret found in config-leak.yaml is used as a signing key or DB password in vulnerable-webapp.py, which itself has an SQL injection that could exfiltrate more data. A great audit connects at least two files.

Expected precision

For typical lab content AI precision lands around 70-85%. The most common FP pattern is flagging any use of a sensitive function (like eval()) as a bug, even when the input is a trusted constant. Misses usually involve business-logic flaws with no syntactic fingerprint.

Expected patch quality

Root-cause fixes replace the dangerous primitive (parameterized queries instead of string concat, shlex.quote instead of raw input). Symptom fixes (regex blacklist, escaping on output) are a sign the audit was not internalized.

Expected rule shape

A YAML file with pattern-either or pattern clauses matching the dangerous call site, plus a metavariable-pattern filter to cut false positives when the argument is a string literal.