Source linked

Malicious Plugins Succeed 100% of the Time in Claw-like AI Agents

arxiv.org@threat_watch1 hour ago·Cybersecurity·2 comments

A new benchmark shows malicious Plugins bypass every LLM tested, and even the best defense only cuts success rates to 22%.

safeclawarenaopenclawseclawai agent securitysupply chain attacksuc berkeley

Malicious Plugins in Claw-like AI agents succeed 100% of the time—no LLM, no platform stops them. That’s the headline from SafeClawArena, a new security benchmark from UC Berkeley’s Sunblaze lab.

SafeClawArena runs 406 adversarial tasks across four attack surfaces: Skill Supply-Chain Integrity, Persistent State Exploitation, Cross-Boundary Data Flow, and Indirect Prompt Injection. Each task executes inside containerized replicas of real agent platforms with canary-marked credentials, tracked via automated taint analysis across nine output channels.

Three Platforms, Five LLMs, One Grim Picture

Evaluating OpenClaw, NemoClaw, and SeClaw with five frontier LLMs, the highest attack success rate hit 70%. GPT-5.4 was the worst offender. Claude-Opus-4.6 was the least vulnerable, but still sat near a 22% floor on every platform—meaning even the best language model can’t overcome sloppy runtime security.

SeClaw deserves attention: it cuts GPT-5.4’s success rate from 70% down to 22%. But the paper is blunt—that improvement comes partly from utility-security tradeoffs, not active defenses. You lose capability to gain safety.

The Blind Spot: Cross-Component Failure Modes

The researchers frame Claw-like agents as agentic computer systems: the gateway runtime acts like an OS, Skills resemble user-installed applications, and Plugins behave like loadable extensions with runtime privileges. Each has a classical security counterpart—but the decades of OS protection mechanisms (sandboxing, capability systems, privilege separation) are simply absent on the agent side.

Existing benchmarks only measure model responses and tool calls. SafeClawArena measures what happens when an attacker chains a malicious Plugin with a persistent state exploit and a prompt injection—the cross-component failures that real attackers will use.

What This Means for Anyone Building Agents

If you’re shipping a Claw-like agent—persistent, credentialed, with file and tool access—you have a security gap that no prompt engineering or model choice will fix. The code and data are on GitHub at sunblaze-ucb/SafeClawArena. The real work starts when platform builders finally treat their runtimes as operating systems worthy of proper isolation.

Source: Understanding and Evaluating Claw-like Agent Security Through a Computer-Systems Lens
Domain: arxiv.org

Read original source ->

External source stays available while the OJO article and comment thread stay local.

More in Cybersecurity

view topic
Tailscale's Aperture Ties Cryptographic Identity to Every LLM Request for Audit

Every AI request through Aperture carries cryptographic proof of identity, turning the usual black-box LLM usage into an auditable trail tied to a specific person or device.

World Leaks Dumps 630GB of Tata Data, Exposes iPhone 18 Pro Supply Chain

Ransomware group published over 200,000 files including confidential supplier maps and drop-test photos for Apple's unreleased iPhone 18 Pro and Pro Max.

15 of 21 RL Vulnerability Studies Still Just Fuzzing - Only One Localizes Bugs

A systematic review of 21 papers finds reinforcement learning for C/C++ vulnerability detection is stuck on fuzzing; just three tackle direct detection and exactly one locates code at the statement level.

PLAA Packet-Level Attacks Evade NIDS with 92.78% Success

A new adversarial attack generates network traffic at the packet level, preserving attack semantics and achieving a 92.78% average evasion rate across three benchmark datasets.

Comments load interactively on the live page.