matrix-handoff
Pass
Audited by Gen Agent Trust Hub on Aug 20, 2026
Risk Level: SAFECREDENTIALS_UNSAFECOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [DATA_EXPOSURE]: The skill is designed to search for and read sensitive agent conversation transcripts from the user's home directory.
- Evidence: The
findClaudeTranscriptandfindCodexTranscriptfunctions inscripts/matrix-handoff.mjsscan~/.claude/projects/and~/.codex/sessions/for.jsonlfiles to include in the handoff bundle. - [COMMAND_EXECUTION]: The script executes several local system commands to package the environment and interact with the Matrix platform.
- Evidence:
scripts/matrix-handoff.mjsusesexecFileto rungit(for file discovery and metadata),tar(for bundling), and thematrixCLI (for uploading and session management). - [REMOTE_CODE_EXECUTION]: The skill generates a shell script locally and executes it on the remote Matrix host to extract the project bundle.
- Evidence: The
buildRemoteExtractScriptfunction inscripts/matrix-handoff.mjsconstructs a multi-line shell command that is passed tomatrix run -- sh -lcon the remote system. - [INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface where a remote agent is instructed to process untrusted data that could contain malicious instructions.
- Ingestion points: Reads all cached and untracked files in the repository (via
git ls-files), local agent transcripts (.jsonl), and an agent-written continuation brief. - Boundary markers: The remote agent is prompted with:
"Continue this task from the Matrix handoff. First read .matrix-handoff/README.md and .matrix-handoff/manifest.json, inspect the working tree, then proceed from the continuation brief."No explicit delimiters or safety warnings are included to prevent the agent from obeying instructions hidden within the brief or the repository files. - Capability inventory: The remote agent session is started with a full shell (
matrix shell new) and has the same capabilities as a standard Matrix OS agent. - Sanitization: The skill uses
isSafeHandoffPathto filter sensitive file names andshellQuoteto sanitize the final prompt string, but it does not sanitize the content of the uploaded files or the transcript for embedded instructions.
Audit Metadata