How to use Claude to audit a forensic template
Claude Code helped find a Velociraptor template flaw after human-guided code auditing. A harmless local proof and private disclosure led to CVE-2026-8795.
shellcode.blog
Original prompt
Shared by the source author. See source notes for context.
As a code review expert, your role will be to carefully examine the code for potential security flaws.
Reach people exploring AI agents
See placements →What happened
A benign local proof supported a privately reported finding that became CVE-2026-8795. Human steering and rejection of low-impact leads were part of the discovery.
Reported by Mohammed Alshehri — Security research blogHow the agent did it
- Alshehri redirected noisy audit results toward collection remapping and other familiar trust boundaries.
- Claude traced untrusted metadata through the template and constructed a benign local proof.
- The researcher verified impact and privately reported the issue and proposed correction.
Where a person stepped in
Alshehri selected trust boundaries, rejected low-impact leads, required a harmless reproducible proof and privately reported the finding.
Mohammed Alshehri used Claude Code to audit Velociraptor and found a template-handling vulnerability later assigned CVE-2026-8795. The initial audit produced noisy, low-impact leads, so he redirected the agent toward workflows he understood, including collection remapping and repacking.
That human choice of trust boundary mattered. The eventual issue involved attacker-controlled collection metadata entering a Windows collection-remapping template without safe handling. In the reported chain, opening an untrusted collection could reach code execution on the analyst's machine rather than remaining data inside the forensic workflow.
Claude analyzed the repository and wrote a Python helper to construct a laboratory collection. It iterated the template and YAML handling until a harmless file-creation check demonstrated the behavior. Alshehri required actual impact rather than accepting a plausible code-review warning, then used the agent's report and proposed correction in private disclosure.
The source describes acceptance by the project and subsequent public disclosure of the CVE. It supports one reproducible finding, not a claim that the agent can reliably audit every project. The useful workflow combines a human-selected attack surface, agent analysis and a benign local proof, followed by maintainer review. Dead ends and researcher steering were part of the discovery.
Prompt
Original prompt
Source: Security research blog · Mohammed Alshehri
Frequently asked questions
Why redirect the initial audit?
Its early findings were noisy and lacked the desired impact.
How was the finding verified?
A laboratory collection caused a harmless file-creation result.
What was the disclosure result?
The source reports project acceptance and public CVE-2026-8795 disclosure.
More details
About this account
Alshehri redirected noisy audit results toward collection remapping and other familiar trust boundaries. Claude traced untrusted metadata through the template and constructed a benign local proof.
Original researcher account opened. The final vulnerability received CVE-2026-8795; the reported discovery time is self-reported and no general autonomous-auditing success rate is established. Original prompt excerpt.
Credit: Mohammed Alshehri — Security research blog
Source published 27 Aug 2026
Source checked 14 Oct 2026
Tools and topics
Claude Code · Velociraptor · Python · Ubuntu
Vulnerability research · Responsible disclosure
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.