Claude Opus Wrote the App and the Dependencies — SAST and SCA Caught Both

This guide uses a synthetic InventoryDesk parts-tracking service written for the walkthrough only. It is not a real product and was never deployed. Everything under app/ and the contents of requirements.txt came from one prompt sent to Claude Opus in Cursor.

The prompt asked for a small internal parts desk with login, searchable lists, CSV reports, document upload, filter presets, and admin endpoints. It deliberately told the model to add whatever Python packages fit a small FastAPI service — and did not ask for security review, pinned versions, or dependency verification.

That open-ended dependency instruction is common in real prompts. Models often produce plausible package names that differ by a hyphen from the canonical PyPI name, and pin versions that were current in training data but carry known CVEs today.

Save the prompt text alongside the archive and requirements.txt you upload, so later readers can see what the model was asked to produce rather than what an engineer added afterwards.

Claude Opus added six packages. Two are the ground truth for the dependency half of this guide.

Pillow==8.0.1 is a real package on a version with many published CVEs. On the captured installation SCA reported thirty-four advisories on that component alone, including CVE-2020-35653 (buffer over-read in PCX decode, fixed in 8.1.0).

django-rest-framework-jwt==1.0.0 is not the canonical PyPI name for the well-known JWT helper — that name is djangorestframework-jwt. SCS flagged the hyphenated name as a possible typosquat of the legitimate library. This is slopsquatting-adjacent behaviour, not proof that the package is malicious, but enough to stop and verify before install.

Request a token with the account credentials and a name you will recognise later. Send the access token as a bearer credential on every subsequent request.

If token creation returns SESSION_LIMIT_REACHED, clear active JWT sessions on the installation before retrying. The local stack in this guide uses a single session slot.

Upload the flat zip archive with languages=PYTHON. Keep every aiTriage flag off so findings stay unverified until you choose otherwise.

On the captured installation the Classic Python engine finished in about two seconds and reported 792 lines, one critical finding, sixteen medium, and a security score of 40.

Grouped by rule, the scan reported one SQL injection, two command injection, eight path manipulation, two deserialization, and one each for weak hashing, weak randomness, hardcoded key, and resource injection.

The clearest SQL injection is string concatenation in the list endpoint — search text, column name, status, and sort parameters are assembled directly into the query.

Command injection appears where admin endpoints call subprocess.run with shell=True on a string built from request parameters.

Create an SCA project and upload the same source archive. Enable scaAnalysis, scsAnalysis, and Hybrid analysis — the sast flag in the API — so DerScanner runs SCA first, then SAST over imports and calls in the uploaded code.

On the captured installation online SBOM generation succeeded with offlineSbomGeneration set to false. The scan progressed through SBOM, SAST, and SCS stages and finished in about two and a half minutes.

Six components were reported. Pillow carried thirty-four SCA issues; python-multipart carried eight; django-rest-framework-jwt carried one SCS issue. The Dependency Tree tab shows Pillow and python-multipart with severity badges on the graph nodes.

Open a Pillow CVE in Detailed Results and switch to Call Trace to see the reachability graph from the project node to pillow 8.0.1. That graph is the UI expression of hybrid SAST+SCA analysis — it helps you decide whether a CVE in a third-party package is relevant before you prioritise an upgrade.

The Dependency Tree tab renders an interactive graph of direct dependencies. Nodes for Pillow and python-multipart carry severity badges — twenty-two critical, nine medium, and three low on Pillow in the captured scan.

Detailed Results on a specific CVE — here CVE-2020-35653 on Pillow — exposes a Call Trace panel below the advisory text. The graph runs from InventoryDesk SCA hybrid guide on the left to pillow 8.0.1 on the right, with the same severity counts on the component node.

Reachability analysis does not replace upgrading vulnerable packages, but it narrows the list you treat as urgent. A CVE with no import path into your code may still matter for compliance, yet it is a different remediation conversation from one whose call trace lands in a route handler.

Select Pillow in Detailed Results, expand CVE-2020-35653, and enable Call Trace. The panel renders an ngx-graph view from the project root to the vulnerable component.

CVSS, EPSS, and alias links remain on the advisory tabs above the graph. The call trace answers a different question — whether this dependency version sits on a path the application actually uses.

Open Detailed Results on the SCA project and select django-rest-framework-jwt. The issue type is Supply Chain Risk, not a CVE.

The SCS panel lists possibleTyposquattingLibraries with djangorestframework-jwt — the name developers expect for the established library. The hyphenated variant is what Claude Opus wrote; nothing in the prompt asked for that specific package.

Treat this as a verification gate, not a verdict. Confirm the name against PyPI, your internal allow-list, and your package manager lockfile before dismissing or accepting the dependency. See the slopsquatting knowledge page for the broader pattern.

SAST answers whether the generated source contains exploitable patterns. Here that is yes — concatenated SQL and shell=True command execution are reachable from normal API parameters.

Hybrid SCA answers whether pinned versions carry published CVEs and whether those CVEs connect to code paths in the project. Pillow 8.0.1 does, and the call trace makes the dependency edge visible in the UI.

SCS answers whether a package name looks wrong relative to known libraries. django-rest-framework-jwt fails that check even if the version string is innocuous.

A team that only runs SAST on the zip would ship vulnerable dependencies without reachability context. A team that only runs SCA without hybrid analysis would see CVE counts but not how those components attach to the codebase. Running SAST plus hybrid SCA is the minimum credible review for AI-generated services with model-chosen requirements.