Case · DiggingBeagle record

A caller-controlled jq filter could terminate or monopolize a ContextForge worker despite a dangerous-function denylist

The ContextForge tool-response transformation path exposed a jsonpath_filter accepted from an authenticated remote caller. The code attempted to reject dangerous jq built-ins using an incomplete blacklist rather than enforcing a bounded expression language or process resource limit.

Public disclosure: Nov 13, 2026

Scope

An authenticated caller can supply response filters through a gateway feature enabled in the deployment.

UnratedAI role: AGAINST AIAssessment method

At a glance

Mechanism and trust boundary

  1. 01

    Starting condition

    An authenticated caller can supply response filters through a gateway feature enabled in the deployment.

    Boundary: precondition and attacker capability

  2. 02

    Trust boundary crossed or evaluated

    IBM identifies halt, halt_error and unbounded-iteration classes as omissions that can reach the jq evaluator and terminate a worker or force prolonged execution. This is not the same path as catastrophic regular expressions in registered JSON Schemas.

    Boundary: mechanism

  3. 03

    Observed consequence

    A syntactically accepted filter can exceed the worker execution budget or terminate its evaluator process, disrupting unrelated agent calls.

    Boundary: observed effect and limitations

Claims & evidence

CLM-NOV15-CF110189-MECHThe ContextForge tool-response transformation path exposed a jsonpath_filter accepted from an authenticated remote caller. The code attempted to reject dangerous jq built-ins using an incomplete blacklist rather than enforcing a bounded expression language or process resource limit. IBM identifies halt, halt_error and unbounded-iteration classes as omissions that can reach the jq evaluator and terminate a worker or force prolonged execution. This is not the same path as catastrophic regular expressions in registered JSON Schemas.supported

Basis: reported finding

Link to claim
CLM-NOV15-CF110189-SCOPECVE-2026-110189 affects 1.0.8 through 1.0.11 and 1.1.0, with IBM directing users to 1.1.1. The bulletin does not establish an actual production outage or arbitrary operating-system execution. The MCP-specific risk is that tool data transformation inherits availability authority over a common gateway process.supported

Basis: reported finding

Link to claim

Implications

A syntactically accepted filter can exceed the worker execution budget or terminate its evaluator process, disrupting unrelated agent calls.

Controls and mitigations

  • Upgrade to ContextForge 1.1.1 or later and sandbox filter execution with strict time/CPU/process containment.
  • Treat filter allowlisting as a structural expression-language problem, not a list of suspicious words alone.

Unknowns and contradictions

  • The MCP-specific risk is that tool data transformation inherits availability authority over a common gateway process.

Sources and citation

Material revision history

  1. Nov 15, 2026 · Canonical change recorded · new in release · revision 250