Case · DiggingBeagle record

Malformed provider responses could crash Agent Router extproc

Unexpected but successful-status AI provider response shapes reached nil-pointer handling.

Public disclosure: Nov 6, 2026

Scope

Agent Router extproc response parsing and shared-process failure domain prior to the 1.2.1 fix.

Assessment profile

Assessment method
Setting
research demonstration
Exploitation
demonstrated
Evidence
Grade B
Remediation
patch available
AI role
AGAINST AI
Basis and provenance

The first-party advisory identifies nil-dereference paths in malformed successful provider responses and a 1.2.1 repair. It does not document an actual provider attack or customer outage. An impact band remains unassigned because the available evidence does not establish a comparable severity score or verified numerical loss for this assessment.

Assessed Nov 8, 2026 using diggingbeagle.assessment/1.

At a glance

Mechanism and trust boundary

  1. 01

    Return a malformed HTTP 2xx provider response with absent required nested values.

  2. 02

    Trigger a nil dereference inside extproc processing.

  3. 03

    Interrupt other requests handled by the same process.

Claims & evidence

CLM-AGENTROUTER-EXTPROC-NIL-RESPONSES-01Several malformed HTTP 2xx responses from supported AI providers could cause nil-pointer dereferences in Agent Router extproc, disrupting AI inference and a co-located MCP proxy process.supported

Basis: reported finding

Link to claim

Implications

Validation of untrusted successful responses matters for availability as well as correctness, especially in shared AI gateways.

Controls and mitigations

  • Upgrade Agent Router extproc to 1.2.1 and verify null or omitted nested fields from each supported provider return controlled errors without panics.
  • Consider isolating the MCP proxy from any single malformed provider response failure so one request does not interrupt unrelated traffic.

Unknowns and contradictions

  • An upstream response must take one of the malformed shapes; the report does not establish a malicious provider attack, customer outage or remote arbitrary code execution.

Sources and citation

Material revision history

  1. Nov 8, 2026 · Canonical change recorded · new in release · revision 229