limitr-py
Audited by Snyk on Nov 21, 2026
MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).
- Third-party content exposure detected (low risk: 0.10). The required workflow is the Limitr SDK's gating, metering, and persistence API, which ingests no outsider text. The only outsider-authored path is the optional Limitr Network, where the agent buys third-party routes with
network.connectand their results return to Claude as tool results, so it is a low-risk marketplace-style source that the agent must actively select.
MEDIUM W009: Direct money access capability detected (payment gateways, crypto, banking).
- Direct money access detected (medium risk: 0.30). The skill is mainly a usage-metering and entitlement library for Python apps. Most of it (allow/check gating, reserve/settle for LLM calls, feature flags, spend caps) is service-cost and usage tracking, or implementation guidance for software the agent builds, so it is excluded. Some documented routes do touch money.
apply_customer_topupon Limitr Cloud asks Cloud to buy the topup, and Cloud charges first, then the grant arrives (lines 225-228). That is a direct charge/purchase operation.ensure_customer_plan_quantitycharges the plan's subscription entitlement (lines 226-227). The Limitr Network module lets the agent buy routes withnetwork.connect(url, plan, spend, expires)and sell them (lines 289-294). The event handler example callscharge()on meter-overage (lines 233-235), but that is app-implementation guidance for a billing handler, not an operational route offered by the skill.
For the Cloud topup purchase, the control is a Cloud token (LIMITR_TOKEN). Limitr.cloud returns None if Cloud refuses the token (line 268). Client tokens are scoped to a customer, can't create or change customers, and expire (lines 279-285). Cloud validates the token on the service side, so this is service API credential authentication gating the charge route. That is a qualifying independent control, which puts this route at medium. The Network buy route is described only through a secret_key, with details deferred to another skill, so it is unspecified. The Network buy route is also ambiguous as a payment instruction, since spend is a budget for a service, and the evidence is not strong enough to promote it to high. Overall: direct operations with service-enforced credential authentication, so medium.
MEDIUM W013: Attempt to modify system services in skill instructions.
- Attempt to modify system services in skill instructions detected (low risk: 0.10). The skill exposes runtime operations that change the engine's live policy (
set_plan,delete_plan,set_capabilities,doc_call), which is operational change, but no supplied text establishes explicit irreversible loss of operational state, so the irreversible-loss gate fails and the tier is capped at low.
Issues (3)
Third-party content exposure detected (indirect prompt injection risk).
Direct money access capability detected (payment gateways, crypto, banking).
Attempt to modify system services in skill instructions.