Case · DiggingBeagle record

An MCP tool schema could carry an expensive regular expression that blocks the shared ContextForge worker

ContextForge accepts JSON Schemas for registered tools, prompts or federated tool surfaces. An actor allowed to register a schema can embed a regular expression with catastrophic backtracking in pattern or patternProperties and cause expensive validation on later arguments.

Public disclosure: Nov 13, 2026

Scope

An authenticated actor has tool/prompt registration rights or can supply a federated schema that reaches the validator.

UnratedAI role: AGAINST AIAssessment method

At a glance

Mechanism and trust boundary

  1. 01

    Starting condition

    An authenticated actor has tool/prompt registration rights or can supply a federated schema that reaches the validator.

    Boundary: precondition and attacker capability

  2. 02

    Trust boundary crossed or evaluated

    The danger appears when the gateway evaluates otherwise legitimate tool arguments against the retained attacker-influenced schema. The tool registration event and subsequent resource exhaustion may be separated in time or occur under different callers.

    Boundary: mechanism

  3. 03

    Observed consequence

    A schema is stored as trusted validation policy despite attacker-selected regex work, letting later tool calls consume worker CPU.

    Boundary: observed effect and limitations

Claims & evidence

CLM-NOV15-CF94407-MECHContextForge accepts JSON Schemas for registered tools, prompts or federated tool surfaces. An actor allowed to register a schema can embed a regular expression with catastrophic backtracking in pattern or patternProperties and cause expensive validation on later arguments. The danger appears when the gateway evaluates otherwise legitimate tool arguments against the retained attacker-influenced schema. The tool registration event and subsequent resource exhaustion may be separated in time or occur under different callers.supported

Basis: reported finding

Link to claim
CLM-NOV15-CF94407-SCOPEIBM reports CVE-2026-94407 for v1.0.0 through v1.0.11 and recommends upgrade to 1.1.0. This establishes a code-level denial-of-service risk, not a confirmed service outage or model-driven misuse. This is a security gap in AI tool metadata admission, not a general inability of the language model to parse a malicious prompt.supported

Basis: reported finding

Link to claim

Implications

A schema is stored as trusted validation policy despite attacker-selected regex work, letting later tool calls consume worker CPU.

Controls and mitigations

  • Enforce schema pattern complexity and execution budgets before accepting new tools.
  • Upgrade to fixed gateway versions and test admission on hostile patterns without allowing unbounded validation of shared worker traffic.

Unknowns and contradictions

  • This is a security gap in AI tool metadata admission, not a general inability of the language model to parse a malicious prompt.

Sources and citation

Material revision history

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