differential-review

name	differential-review
description	Security-focused git diff analysis to detect regressions, new vulnerabilities, and unsafe changes in code reviews
license	MIT
metadata	{"category":"code-security","locale":"en","phase":"v1"}

What this skill does

Analyzes git diffs, pull request changes, and commit ranges for security regressions. Unlike a full codebase audit, this skill focuses exclusively on what changed: additions that introduce new vulnerabilities (hardcoded secrets, injection sinks, unsafe deserialization), and removals that eliminate security controls (auth checks, input validation, rate limiting). It classifies changed files by risk tier, runs targeted pattern checks on additions and deletions, cross-references dependency changes, and produces a prioritized findings list with severity.

When to use

• During pull request review when you need a security-focused lens on the diff specifically
• When a commit range is suspected of introducing a regression (e.g., after an incident or pentest finding)
• As a lightweight pre-merge gate before running a full secure-code-review
• When reviewing dependency bumps that may silently alter security behavior
• When a developer says "I only changed X" and you need to verify that claim from a security perspective

Prerequisites

• Git access to the repository (local clone or remote diff URL)
• git CLI available in the environment
• (Optional) semgrep for automated pattern matching on additions
• Read access to dependency manifests (package.json, go.mod, requirements.txt, Cargo.toml, etc.)

Inputs

Item	Description	Example
DIFF_SOURCE	Source of the diff to review	git diff main..feature/login, PR URL, patch file
BASE_REF	Base branch or commit	main, v2.3.1, abc1234
HEAD_REF	Target branch or commit	feature/auth-refactor, HEAD, def5678
REPO_PATH	Path to the local repository	/home/user/myapp
LANGUAGE	Primary language(s) in the diff	Python, TypeScript, Go

Workflow

Step 1: Obtain the diff

Choose the appropriate method based on DIFF_SOURCE:

# Commit range
git diff <BASE_REF>..<HEAD_REF>

# Staged changes only
git diff --cached

# Specific files in a range
git diff <BASE_REF>..<HEAD_REF> -- path/to/file

# Full diff with context (recommended for security review)
git diff -U5 <BASE_REF>..<HEAD_REF>

# List only changed files for triage
git diff --name-only <BASE_REF>..<HEAD_REF>

# Include commit messages for context
git log --oneline <BASE_REF>..<HEAD_REF>

If working from a PR diff URL, download the raw .diff or .patch and treat it as a local file.

Step 2: Classify changed files by risk tier

Before reading line-by-line, triage which files deserve the most attention.

High risk — review every line of additions and deletions:

Signal	Examples
Authentication / session	auth.py, login.ts, session_manager.go, jwt_util.rb
Authorization / access control	permissions.js, rbac.go, policy.py, middleware/auth*
Cryptography	crypto/, cipher.py, hash_util.ts, tls_config.go
Input parsing / deserialization	parser.py, deserialize.ts, xml_handler.go, yaml_loader.rb
SQL / query construction	db.py, queries.ts, repository.go, dao/*.java
Secret / config handling	.env*, config.yaml, secrets.ts, vault_client.go
Dependency manifests	package.json, go.mod, requirements.txt, Cargo.toml

Medium risk — review additions and significant deletions:

• HTTP route handlers and controllers
• File upload / download handlers
• External API client code
• Logging and error handling modules

Low risk — spot-check only:

• UI/template changes with no server-side logic
• Documentation and comments
• Pure test files (unless they reveal implementation details)
• Build scripts with no runtime impact

Step 3: Audit additions for security anti-patterns

For each + line in high- and medium-risk files, check for the following. A match is a finding candidate.

Secrets and credentials

# Patterns to flag in additions
(?i)(password|passwd|secret|api_key|apikey|token|auth_token)\s*=\s*['"][^'"]{6,}['"]
(?i)(aws_access_key|aws_secret|private_key)\s*=
-----BEGIN (RSA|EC|OPENSSH) PRIVATE KEY-----
(?i)bearer\s+[a-z0-9\-_]{20,}

Injection sinks

# SQL concatenation (any language)
query = "SELECT.*" + user_input
execute("INSERT INTO.*" + var)
db.query(`SELECT * FROM ${req.body`)

# Command injection
os.system(user_input)
subprocess.call(cmd, shell=True)
exec(user_controlled)
child_process.exec(req.query

# Template injection
render_template_string(user_input)   # Jinja2
new Function(userCode)               # JS
eval(expression)                     # any language

Unsafe deserialization

pickle.loads(data)           # Python
yaml.load(data)              # Python — safe only with Loader=yaml.SafeLoader
ObjectInputStream            # Java
JSON.parse + eval            # JS with execution
BinaryFormatter.Deserialize  # .NET

Path traversal

open(user_input)
os.path.join(base, user_input)   # without normalization + prefix check
fs.readFile(req.params.filename)

Cryptographic weaknesses

MD5 / SHA1 for password hashing or HMAC
DES / 3DES / RC4 / ECB mode
random() / Math.random() for tokens or nonces
hardcoded IV or salt

If semgrep is available, run targeted rulesets against the patch file:

semgrep --config=p/secrets --config=p/owasp-top-ten <patch_file_or_dir>
# Or against only changed files:
git diff --name-only <BASE>..<HEAD> | xargs semgrep --config=p/security-audit

Step 4: Audit deletions for removed security controls

For each - line in high-risk files, check whether the removal eliminates a protection. A protection removal with no replacement is a finding.

Categories to flag:

Removed pattern	Risk
Authentication middleware / decorator removed from a route	Unauthenticated access
Input validation / sanitization call removed	Injection or XSS vector opened
Rate limiting or throttle removed	Brute force / DoS exposure
CSRF token check removed	CSRF vulnerability
Authorization assertion removed (require_permission, can?, hasRole)	Privilege escalation
TLS/cert verification disabled (verify=False, InsecureSkipVerify)	MitM exposure
Security header removed from response	Browser-side protection lost
Audit log call removed	Loss of forensic visibility
Timeout or size limit removed	DoS / resource exhaustion

When a protection is removed, check whether it is reinstated elsewhere in the diff (e.g., moved to a different layer). If not replaced, record as a finding.

Step 5: Review dependency changes

Examine additions and removals in dependency manifests.

# Show only manifest changes
git diff <BASE>..<HEAD> -- package.json go.mod requirements.txt Cargo.toml pom.xml

# For npm: check advisories on newly added packages
npm audit --audit-level=moderate   # after npm install

# For Python
pip-audit   # or: safety check -r requirements.txt

# For Go
govulncheck ./...

Flag:

• New direct dependencies with known CVEs
• Version downgrades (could reintroduce fixed vulnerabilities)
• Pinned versions removed (hash pinning, ==x.y.z replaced with >=x.y)
• New transitive dependencies that are known malicious or typosquatted

Step 6: Produce findings

For each identified issue, record:

### [SEVERITY] <Short title>

- **File**: `path/to/file.py:42`
- **Commit / hunk**: `abc1234` or diff hunk reference
- **Type**: Addition of vulnerable pattern | Removal of security control | Dependency risk
- **Description**: What was introduced or removed and why it is a security concern.
- **Impact**: What an attacker could achieve if exploited.
- **Remediation**: Specific fix — replace `eval(x)` with safe alternative, restore the auth check, pin the dependency version, etc.
- **References**: CWE-XX, CVE-YYYY-NNNNN (if applicable)

Severity guide for differential review:

Severity	Criteria
CRITICAL	Auth bypass, RCE sink, credential leak committed to repo
HIGH	SQL/command injection, unsafe deserialization, removed auth check on sensitive route
MEDIUM	Path traversal, weak crypto introduced, rate limiting removed
LOW	Debug logging of sensitive data, security header removed, soft crypto downgrade
INFO	Style issue with minor security implication, dependency without known CVE but unmaintained

Done when

• All changed files are classified by risk tier
• Every addition in high- and medium-risk files has been scanned for the anti-patterns in Step 3
• Every deletion in high-risk files has been checked for removed security controls (Step 4)
• Dependency manifest changes are reviewed and advisory-checked (Step 5)
• All findings have a severity, file reference, description, impact, and remediation
• A brief summary is produced: N files changed, X high-risk, Y findings (Z CRITICAL/HIGH, W MEDIUM/LOW)

Failure modes

Issue	Cause	Solution
Diff is too large to review manually	Large PR with hundreds of changed files	Triage by risk tier first; focus on CRITICAL and HIGH risk files; flag the PR for splitting
Diff context is insufficient to judge a deletion	-U0 diff with no surrounding lines	Re-run git diff -U10 or read the full function from the source tree
False positive on secret pattern	Test fixtures or placeholder strings	Confirm the value is not used in production config; note as INFO if test-only
Semgrep not available	Tooling gap	Proceed with manual pattern check; note in report that automated scan was skipped
Dependency manifest not present in diff	Only lock file changed	Review the lock file diff for transitive dependency changes; run npm audit / pip-audit locally
Renamed file appears as delete + add	Git rename detection off	Re-run git diff -M to detect renames; treat as a single change, not a removal

Notes

• This skill reviews changes, not the full codebase state. A vulnerability that existed before the diff and was not touched is out of scope — use secure-code-review for that.
• This skill pairs well with secure-code-review (full domain-based audit) and secret-scan (credential detection in diffs).
• A finding of type "removal of security control" is often more urgent than an addition finding, because it may silently expose existing functionality to attack.
• When reviewing auth-related deletions, check the commit message and linked issue for intent — accidental removals and intentional refactors look identical in a diff.
• Keep findings focused on the diff. Avoid reporting pre-existing issues unless they are directly worsened by the change.

---

Adapted from Trail of Bits via awesome-agent-skills.