ai-agent-security

topic6 events

papersTODAY 04:00 UTC

arXiv Paper Describes Persistent Memory Poisoning Attack on Harness-Based LLM Agents

A new arXiv preprint examines how harness-based LLM agents, which combine memory, tool use, and runtime control, can be compromised through stored malicious instructions. The authors argue that once such instructions enter an agent's persistent memory, they can continue to influence later behavior, creating security and privacy exposure. The work frames memory poisoning as a distinct risk for agent architectures that retain context across sessions.

safety/agents/security/memory

papersTODAY 04:00 UTC

arXiv Paper Proposes Task-Based Permission Scoping for AI Agents

A new arXiv preprint examines how enterprise AI agents are typically given static credentials at deployment that mirror the full set of permissions an employee role could hold. The authors argue this approach, inherited from role-based access control, grants agents far more access than any single task requires. They evaluate an alternative architecture that scopes an agent's permissions to the specific task it is performing.

agents/security/access-control/arxiv

papersTODAY 04:00 UTC

arXiv paper proposes runtime authorization for resources acquired by AI agents

A new arXiv preprint titled "AcquireBound" examines how autonomous AI agents gain new authority by acquiring compute, credentials, accounts, services, and other agents during a task. The authors argue that existing payment, budget, OAuth, mandate, and fulfillment checks verify transaction conditions but do not resolve whether the accumulated authority itself should be permitted. The work proposes runtime authorization as a way to bound the resources an agent may acquire while operating.

agents/security/authorization/arxiv

papersTODAY 04:00 UTC

SkillAtlas: An Attack Trace Library for Agent Skills

Researchers present a library of attack traces aimed at reusable skills for language-model agents. The work argues that risks in agent skills surface through model decisions, user context, tool calls, and execution feedback rather than through fixed signatures or a single sandboxed run, which limits existing static and dynamic analysis methods. The library is intended to help catalog and study these behaviors.

agents/safety/research/llm-agents

papersSEP 10 04:00 UTC

AgentHijack: Visual Patch Attacks on Multimodal Computer-Use Agents

A new arXiv paper introduces an end-to-end evaluation framework for testing whether a locally placed visual patch can hijack multimodal computer-use agents into executing attacker-chosen commands. Rather than stopping at model-level manipulation, the study checks whether such image-triggered injections lead to verifiable consequences in the agent's operating environment. The work adds to a growing body of research on the security risks of AI agents that control graphical interfaces.

agents/security/adversarial/benchmark

industrySEP 9 13:00 UTC

Sequoia invests again in Cymphony, a security platform for AI agent identities

Venture firm Sequoia has made a second investment in Cymphony, a startup addressing security risks introduced by AI agents in corporate settings. The platform lets security teams see employees, AI agents, and other nonhuman identities in one place, along with the systems and sensitive data each can reach. The funding comes as enterprises grapple with how to govern autonomous agents that hold credentials and access.

funding/agents/security/enterprise