Releases·Tool Release
Authorization gets checked before the tool action—what if it changes before execution?
A developer released AuthDrift, an MIT-licensed Python harness that deterministically tests whether long-running agent workflows re-verify authority before committing side effects. It pauses a workflow after it observes authorization, changes that authority state, resumes the workflow, and checks whether the effect still committed. The tool targets a gap in agent security testing where authorization checked early can become stale before execution.
8 engagement1 sourceFri, Sep 18, 2026, 02:54 AM
AI Overview
- Cheerstopriya released AuthDrift, an MIT-licensed Python harness that pauses workflows, changes or revokes authority, and tests whether the effect still commits
- Refund fixture: vulnerable version reports 20/20 REVOCATION_ESCAPE (exit 1), corrected version 0/20 (exit 0); demos need no LLM credential
Related
Social·Community
An AI Agent Can Have Permission and Still Be Wrong to Proceed
A Reddit post argues that agent permission models answer only a static question — whether an agent is allowed to perform an action — while ignoring whether authorization still holds given changes since it was granted. The author contends each individual permission (reading customer records, updating a CRM, calling an internal API, sending a payment request) can be individually valid yet wrong for the agent to proceed with mid-workflow. The post is a conceptual argument for time-aware authorization in agent systems rather than a report of a new tool or research result.
A Reddit post argues that agent permission models answer only a static question — whether an agent is allowed to perform an action — while ignoring whether authorization still holds given changes since it was granted. The author contends each individual permission (reading customer records, updating a CRM, calling an internal API, sending a payment request) can be individually valid yet wrong for the agent to proceed with mid-workflow. The post is a conceptual argument for time-aware authorization in agent systems rather than a report of a new tool or research result.
13 engagement1 sourceredditSat, Sep 5, 2026, 11:08 PM
Products·launched
FreshCtx — pre-action validation boundary for stale-state agent workflows in LangGraph
FreshCtx is an Apache-2.0 Python library that adds a validation step immediately before protected tools or nodes execute in agent workflows. It lets developers declare which external evidence a tool depends on so resumed LangGraph runs fail or re-check when that evidence is stale. It targets agent builders whose LLM decisions were sound when made but act on changed external state by execution time.
FreshCtx is an Apache-2.0 Python library that adds a validation step immediately before protected tools or nodes execute in agent workflows. It lets developers declare which external evidence a tool depends on so resumed LangGraph runs fail or re-check when that evidence is stale. It targets agent builders whose LLM decisions were sound when made but act on changed external state by execution time.
14 engagement1 sourceredditWed, Sep 9, 2026, 03:16 AM
Social·Community
Agent approvals should bind scope, expiry, and replay rules, not become standing permissions
A Reddit post argues that agent approvals must stay valid only as long as the agent's context, using a staging-deploy timeout scenario where an agent resumes thirty minutes later with stale approval and plan. It proposes binding grants to a specific action, resource, arguments, expiry, replay rule, and audit record.
A Reddit post argues that agent approvals must stay valid only as long as the agent's context, using a staging-deploy timeout scenario where an agent resumes thirty minutes later with stale approval and plan. It proposes binding grants to a specific action, resource, arguments, expiry, replay rule, and audit record.
27 engagement2 sourcesredditTue, Sep 8, 2026, 01:48 PM
Developments·Paper
arXiv paper defines cross-substrate authority gap in multi-agent systems, shows execution-time checks block unsafe actions
A new arXiv paper describes the cross-substrate authority gap in multi-agent systems: authorization-relevant state can live in runtimes, registries, or approval services outside the planner-visible workspace and memory. This means identical final files may require opposite safe actions, complicating safe agent harness design.
A new arXiv paper describes the cross-substrate authority gap in multi-agent systems: authorization-relevant state can live in runtimes, registries, or approval services outside the planner-visible workspace and memory. This means identical final files may require opposite safe actions, complicating safe agent harness design.
0 engagement1 sourcearxivTue, Sep 8, 2026, 09:11 AM
Products·building
Maetra: approval-bound authorization for consequential agent actions
Maetra is an authorization layer for LLM agents that binds human approvals to exact tool payloads. When an agent requests a consequential action like a refund, a reviewer approves a specific envelope (agent, tool, account, amount, policy version, executor); execution is rejected if the payload drifts or if the capability is replayed. It targets builders who need verifiable review-to-execution integrity for agent-driven actions.
Maetra is an authorization layer for LLM agents that binds human approvals to exact tool payloads. When an agent requests a consequential action like a refund, a reviewer approves a specific envelope (agent, tool, account, amount, policy version, executor); execution is rejected if the payload drifts or if the capability is replayed. It targets builders who need verifiable review-to-execution integrity for agent-driven actions.
1 engagement1 sourceredditTue, Sep 15, 2026, 05:00 PM