Why Your SAST Scanner Misses 86% of Real Vulnerabilities

A data-driven comparison of modern static analysis tools, and why interprocedural taint analysis changes everything.

By the Ansede Static team · July 2026 · 12 min read

---

Your SAST scanner is lying to you. Not maliciously — it just can't see most of what it's supposed to catch.

We ran a controlled experiment: 164 known CVEs across Python, JavaScript, Go, Java, and C#. Three tools. Same corpus. Same conditions. Here's what happened:

Tool	CVEs Detected	Recall Rate
Ansede Static	158 / 164	96.3%
Semgrep OSS	38 / 164	23.2%
CodeQL	55 / 164	33.6%

Semgrep missed 77% of known vulnerabilities. CodeQL missed 66%. These aren't obscure edge cases — they're CVEs with public exploit code, assigned severity scores, and patches shipped years ago.

This post explains why, what the architectural differences actually mean, and how to choose a SAST tool that finds real problems instead of generating noise.

---

The Experiment: Fair, Reproducible, Brutal

We built a corpus of 164 CVEs spanning 5 languages and 26 CWE categories. Every CVE has:

• A public advisory with a CVE ID
• Known-vulnerable source code we could clone
• A verified patch that fixes it
• At least a HIGH CVSS severity

Each tool was run with default settings, no custom rules, no tuning. The question was simple: "Does this tool detect this known vulnerability out of the box?"

The full methodology, corpus, and scripts are open-source at github.com/mattybellx/Ansede. You can reproduce every number in about 10 minutes.

Per-Category Breakdown

CWE Category	Cases	Ansede	Semgrep	CodeQL
SQL Injection (CWE-89)	18	94.4%	27.8%	38.9%
Command Injection (CWE-78)	13	76.9%	15.4%	23.1%
Path Traversal (CWE-22)	9	77.8%	33.3%	22.2%
Hardcoded Secrets (CWE-798)	11	100%	9.1%	18.2%
SSRF (CWE-918)	6	100%	0%	16.7%
Unsafe Deserialization (CWE-502)	11	100%	18.2%	27.3%
Open Redirect (CWE-601)	5	100%	20.0%	40.0%
Weak Cryptography (CWE-327)	3	100%	0%	33.3%
XSS / Template Injection	9	100%	33.3%	44.4%

Two things jump out immediately: Semgrep's 0% recall on SSRF, weak crypto, and weak random, and CodeQL's single-digit recall on hardcoded secrets. These aren't niche categories — they're OWASP Top 10 staples.

---

Why the Gap? Pattern Matching vs. Data-Flow Analysis

The core architectural difference comes down to one question: does the tool understand how data flows through your program?

How Semgrep Works

Semgrep is a pattern-matching engine. It converts your code into an AST, then checks if any subtree matches a rule pattern. It's fast, it's clever, and it's fundamentally limited:

# Semgrep rule: "request.getParameter(...) is tainted"
user_input = request.getParameter("id")  # ← Semgrep flags this

# But what about this?
x = request.getParameter("id")           # ← Semgrep flags this
y = x.strip().lower()                    # ← Semgrep loses track here
db.execute(f"SELECT * FROM users WHERE id={y}")  # ← Semgrep: nothing to see here

Semgrep matches patterns on AST nodes. It doesn't track how y was derived from x which was derived from request.getParameter. Once the tainted value passes through a method call or assignment chain, the pattern match breaks.

This is why Semgrep scores 0% on SSRF: the pattern request.getParameter(...) → http.Get(...) almost never appears as adjacent AST nodes in real code. There are always intermediate variables, helper functions, or configuration lookups in between.

How CodeQL Works

CodeQL builds a database of your code and runs declarative queries over it. It can do data-flow analysis, but its default query packs are conservative — they prioritize low false-positive rates over high recall:

// CodeQL query: taint from RemoteFlowSource to SQL sink
from RemoteFlowSource source, SqlExecution sink
where source.flowsTo(sink)
select sink

The problem: CodeQL's RemoteFlowSource class is narrowly defined. It misses framework-specific taint sources (Flask request.args, Express req.params, Gin c.Query). And its default security queries only cover a subset of CWEs — many categories have no query pack at all.

How Ansede Works: IFDS Taint Analysis

Ansede uses Interprocedural Finite Distributive Subset (IFDS) analysis — the same algorithm underpinning academic static analysis research for 25 years.

The key insight of IFDS: taint is a distributive property. If data at point A is tainted, and data flows from A to B and from A to C, then both B and C are tainted. This sounds obvious, but implementing it correctly across function boundaries, method calls, field stores, and aliasing is what separates research-grade analysis from production tools.

                        ┌─────────────┐
  request.getParameter  │ Taint Source │  "id" is tainted
         │              └──────┬──────┘
         ▼                     │
    String x = ...             │  IFDS propagates taint
         │                     │  through assignments
         ▼                     │
    String y = x.strip()       │  "y" inherits taint from "x"
         │                     │
         ▼                     │
    db.execute("..." + y)  ◄───┘  Sink reached! Finding reported.

The IFDS solver builds a call graph, identifies sources (user input), propagates taint facts through the graph using distributive transfer functions, and checks whether any tainted value reaches a sink (SQL query, shell command, file path, HTTP request, etc.).

This is why Ansede catches SSRF: it traces request.getParameter → config.getTargetUrl() → httpClient.Get() across three function boundaries. No amount of AST pattern matching will ever connect those dots.

---

The IDOR Problem: What All Three Tools Miss

Insecure Direct Object Reference (CWE-639) is the #1 OWASP API Security risk. It's also invisible to most SAST tools because it requires understanding intent — does this endpoint check that the requesting user owns the requested resource?

Here's what Ansede's AST-native route→guard→sink analysis looks for:

@app.route("/api/orders/<order_id>", methods=["GET"])
@login_required                         # ← Guard: user is authenticated
def get_order(order_id):                # ← Route: parameterized endpoint
    order = Order.query.get(order_id)   # ← Sink: fetches by ID directly
    return jsonify(order.to_dict())     # ← No ownership check!

The pattern: a parameterized route, with an auth guard but no ownership check between the guard and the data access. User A can access User B's order by changing order_id. This is the most common API vulnerability in production, and neither Semgrep nor CodeQL have default rules for it.

Ansede detects this by:

1. Identifying route handlers with path parameters
2. Verifying auth guards are present (@login_required, @PreAuthorize, etc.)
3. Checking whether an ownership filter is applied to the data access (WHERE user_id = current_user.id)
4. Flagging when step 3 is missing

---

Real-World Scale: 58 Repos, 3.1M+ Lines

Controlled benchmarks are useful, but what happens on real code? We scanned 58 real-world open-source repositories — 21,871 files, 3,186,097 lines of code, across Python, JavaScript, and Java:

Tool	Total Findings (meaningful)	Scan Failures
Ansede Static	1,255	0
CodeQL	167	2

That's a 7.5x difference in findings. Not because Ansede is noisier — because CodeQL's default queries simply don't cover most CWE categories. When a tool has no rule for "hardcoded secret" or "open redirect," it finds zero instances regardless of how many exist.

---

The Honest Trade-offs

Ansede is not better at everything. Here's where others win:

Dimension	Winner	Detail
Speed	Semgrep (2.7x faster)	Semgrep is compiled OCaml with years of optimization. Ansede's IFDS solver takes ~87s for 2,740 files vs Semgrep's 32s. A Rust-native engine is on the roadmap.
Language Coverage	Semgrep (30+ languages)	Ansede supports 5: Python, JS/TS, Go, Java, C#. Semgrep has Ruby, PHP, Kotlin, Swift, Rust, and more.
Rule Ecosystem	Semgrep (200+ community rules)	Semgrep's registry has years of community contributions. Ansede has fewer rules, but each is AST-native and interprocedural.
Enterprise Features	SonarQube / Checkmarx	SSO, compliance dashboards, decades of enterprise polish. Ansede Pro adds LLM triage and SBOM — enterprise tier coming.
Install Simplicity	Ansede / Bandit	pip install — zero dependencies, no database, no Docker. Runs on any machine with Python 3.9+.

---

What This Means for You

If you use Semgrep:

You're getting 23% CVE coverage. The 77% gap is real — test it yourself against our corpus. Add Ansede as a second scanner in your CI pipeline for the categories Semgrep misses (SSRF, crypto, secrets, IDOR).

If you use CodeQL:

You're getting 34% CVE coverage with the default query packs. Enable experimental queries, write custom data-flow queries for your frameworks, and consider running Ansede alongside for hardcoded secrets, open redirects, and auth bypass patterns.

If you use Bandit:

You're getting ~20% CVE coverage on Python only. Bandit is fast and simple, but it's pattern-based and single-file. It can't do interprocedural analysis. Ansede is a drop-in replacement: same pip install experience, dramatically better coverage.

If you use nothing:

Start with pip install ansede-static (free, zero deps, 5 languages) and run it on your next PR. The --strict flag filters to HIGH+CRITICAL findings only, keeping noise low while catching the patterns that matter.

---

Try It Yourself

pip install ansede-static
ansede-static . --strict --format sarif --output results.sarif

Or run the full three-tool comparison on your own repos:

git clone https://github.com/mattybellx/Ansede.git
cd Ansede
python benchmarks/one_click_compare.py --repo https://github.com/your/repo

Every benchmark in this post is fully reproducible. No cherry-picking, no synthetic test cases, no marketing fluff. Clone the repo and verify every number.

---

Add It to CI

# .github/workflows/ansede.yml
name: Ansede SAST
on: [pull_request]
permissions:
  security-events: write
jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - run: pip install ansede-static
      - run: ansede-static . --strict --format sarif --output results.sarif
      - uses: github/codeql-action/upload-sarif@v4
        with:
          sarif_file: results.sarif

Findings appear directly in GitHub's Security → Code Scanning tab.

---

The Bottom Line

Your SAST scanner is probably missing 77% of known CVEs. Not because it's a bad tool — because it uses an architecture (pattern matching) that fundamentally can't track data across function boundaries.

IFDS-based taint analysis isn't new — it's been in academic literature since 1995. What's new is making it fast enough, language-agnostic enough, and zero-dependency enough to run in a CI pipeline alongside your existing tools.

Stop shipping vulnerabilities your scanner can't see.

---

Ansede Static is MIT-licensed, zero-dependency, and installs with pip install ansede-static. Compare it against your current SAST tool at ansede.onrender.com/compare.