Weekly Trending

A cross-provider AI-agent auditing tool for policy, security, reliability, and business-goal adherence, with CLI, agent plugins, and report outputs.

FounderOS is an open-source command centre for a one-person business. It brings communications, a client funnel, social, finances, knowledge, and named AI agents into one operating dashboard.

SkillJack examines persistent skill backdoors in self-evolving agents, where a poisoned lesson can remain in learned behaviour after the original trace is removed.

UniWorld-Design introduces layer-native image models that generate and decompose images into structured semantic RGBA layers for composition and instruction-driven editing.

GROVE gives persistent visual memory separate layers for immediate observations, episodes, and recurring patterns, each with different retrieval behaviour.

This benchmark examines privacy leakage, impersonation risk, and defenses when persona skills turn behavioural traces and voice into reusable agent assets.

ABSeeker uses backtracked evidence credit so a search agent can identify abandoned research steps that were still necessary to produce its final answer.

ToolArtist frames creative generation as an agentic workflow that combines search, reasoning, and image generation under one policy.

The Personalization Mirage studies persistent-memory models and finds them prone to fabricating user attributes; their confidence is not a reliable signal of correctness.

Agent Against Agent studies red-teaming agents that retain and transfer attack strategies, turning learned adversarial behaviour into a security object.

A machine-checked conformance contract for checkpoint, interrupt, and resume semantics in workflow persistence layers, covering exactly-once effects, checkpoint validity, fork determinism, and crash recovery.

UI-Venus-2 treats interaction with real application interfaces as training material for computer-use agents, using action traces and verification against app worlds to improve behaviour.