SQLite Critical CVEs or LLM Slop?

TL;DR Highlight

LLM-fabricated SQLite CVEs were rated Critical by NVD despite citing nonexistent code

Who Should Read

Security or backend developers who auto-open tickets from CVE scores or run AI agents for vulnerability triage

Core Mechanics

Fake CVEs got rated Critical
- A newly created GitHub repo published a batch of SQLite vulnerability advisories.
- NVD flagged them Critical and CISA's ADP agreed.
The cited code didn't exist
- CVE-2026-51302 named exprComputeOperands(), which wasn't in SQLite 3.41 and was added in mid-2025.
- CVE-2026-51296 cited json.c lines 3555 and 3575, but the file in 3.41.0 is only 2706 lines long.
The patch and PoC were fake too
- For CVE-2026-51303, a diff of 3.51.2 vs 3.51.3 showed no changes to src/expr.c.
- PoCs failed at the parser or ran normally, with no crash.
54 of 55 advisories were fabricated
- The team audited 55 advisories from the same account.
- 54 were completely fabricated and 1 had a real bug wrapped in unverified CVE metadata.
Nothing in the pipeline demanded proof

Community Reactions

Opinion that the signal-to-noise ratio drops, making real CVEs harder to find
- The same commenter noted LLMs are also finding legit CVEs and attackers are likely exploiting that.
Surprise that CVEs aren't verified by anyone
- Commenters without security backgrounds assumed an authority reproduced every CVE.
- Some asked why submissions don't require repro steps like normal bug reports.
A security engineer's experience with their own NVD/GHSA lookup tool
- The CVE count for their products was astronomical compared to the versions they actually run.
Point that the analysis post itself looks AI-written
- One commenter ran it through GPTZero and got a 71% chance of AI.
- They argued that debunking slop with slop is a problem.

How to Apply

Add an existence check before auto-creating CVE tickets
- Today, find any rule that opens tickets from CVSS alone and add a hold step.
- Check the vendor security page (e.g. sqlite.org/cves.html) and for linked commits or PRs first.
Build a slop-CVE red-flag checklist
- Include nonexistent functions, line numbers past EOF, empty CPE fields, and version-range contradictions.
- Route to manual review if any one hits.
Add a code-existence step to AI triage agents
- Before proposing a patch, make the agent confirm the function and lines exist at the target version tag.
- If it fails, return 'insufficient evidence' instead of a patch.
Reproduce PoCs in isolation
- Build the target version in a Docker container with AddressSanitizer on and run the PoC verbatim.
- If nothing crashes, lower the priority.

Try It

copy & run

# Paste into Claude or ChatGPT as-is (checks whether a CVE advisory may be fabricated)

You are a security triage assistant. Check whether the advisory below may be fabricated. Answer 'cannot verify' for anything uncertain and do not guess.

[Advisory]
{{full advisory text}}

[Target software and version]
{{product and version}}

Mark each item as pass / suspicious / cannot verify and give a one-line reason.
1. Does the named function exist in that version?
2. Are the cited line numbers within the file length?
3. Does the vendor security page list the same issue?
4. Do the references include a commit hash or PR?
5. Are CPE and version ranges consistent with the narrative?
6. Is the PoC syntactically valid?

Finish with up to 3 things a human must verify directly.

Next step: if the result is 'suspicious', clone the version tag and grep for the function name yourself.

Terminology

CVEA unique ID assigned to a publicly disclosed security vulnerability. Think of it as a vulnerability's ID number.

CVSSA 0-10 scoring scheme for vulnerability severity. Higher means more dangerous.

PoCProof-of-concept code or input that shows the vulnerability actually occurs. Like repro steps in a bug report.

UAF (use-after-free)A bug where already-freed memory is used again. Like mailing a package to an address you've already moved out of.

AddressSanitizer (ASan)A tool that catches memory errors at runtime. Enable it at compile time and it flags bad memory accesses immediately.

CPEA standard identifier naming a product and version. Scanners use it to decide which software is affected.