Free AI Safety Workshop — in collaboration with explainx.aiSave your seat
AgentbeamBeta
AI agent vulnerability · OpenClaw

GHSA-w5c7-9qqw-6645: OpenClaw inter-session prompts could be treated as direct user instructions

highCVSS 4.0 7.1

Summary Inter-session messages sent via sessionssend could be interpreted as direct end-user instructions because they were persisted as role: "user" without provenance metadata. Affected Packages / Versions - Package: openclaw (npm) - Affected versions: = 2026.2.13) Impact A delegated or internal session could inject instructions into another session that appeared equivalent to externally-originated user input.

Published Feb 18, 2026

Affected packages

openclaw

Ecosystem
npm
Affected
< 2026.2.13
Fixed in
2026.2.13
Check
npm ls openclaw

Advisory details

From the GitHub Advisory Database (CC BY 4.0).

Summary

Inter-session messages sent via sessions_send could be interpreted as direct end-user instructions because they were persisted as role: "user" without provenance metadata.

Affected Packages / Versions

  • Package: openclaw (npm)
  • Affected versions: <= 2026.2.12 (i.e. < 2026.2.13)
  • Fixed in: 2026.2.13 (patched versions >= 2026.2.13)

Impact

A delegated or internal session could inject instructions into another session that appeared equivalent to externally-originated user input.

This is an instruction-provenance confusion issue (confused-deputy style), which can lead to unintended privileged behavior in workflows that trust role: "user" as a sole authority signal.

Technical details

Before the fix, routed inter-session prompts were stored as regular user turns without a verifiable source marker.

As a result, downstream workers and transcript readers could not distinguish:

  • External user input
  • Internal inter-session routed input

Fix

OpenClaw now carries explicit input provenance end-to-end for routed prompts.

Key changes:

  • Added structured provenance model (inputProvenance) with kind values including inter_session.
  • sessions_send and agent-to-agent steps now set inter-session provenance when invoking target runs.
  • Provenance is persisted on user messages as message.provenance.kind = "inter_session" (role remains user for provider compatibility).
  • Transcript readers and memory helpers were updated to respect provenance and avoid treating inter-session prompts as external user-originated input.
  • Runtime context rebuilding now annotates inter-session turns with an explicit in-memory marker ([Inter-session message]) for clearer model-side disambiguation.
  • Regression tests were added for transcript parsing, session tools flow, runner sanitization, and memory hook behavior.

Fix Commit(s)

  • 85409e401b6586f83954cb53552395d7aab04797

Workarounds

If immediate upgrade is not possible:

  • Disable or restrict sessions_send in affected environments.
  • Do not use role alone as an authority boundary; require provenance-aware checks in orchestration logic.

Credit

Reported by @anbecker.

Thanks @anbecker for reporting.

Credited to anbecker.

References

How this page is made: the data panel comes from the GitHub Advisory Database, CISA’s KEV catalog, the NVD and FIRST EPSS, and updates twice a day. No analysis has been written for this entry yet. Methodology

More OpenClaw vulnerabilities

All OpenClaw entries