We just released MalwareBench, a new eval for static malware reverse engineering.
We set out to answer a deceptively simple question: can frontier models actually reverse-engineer malware, not just label it, but recover the concrete indicators (C2 endpoints, dropped paths, registry keys, mutexes, config IDs) that a professional human analyst would pull from a sample? 🧵
11/ So, can frontier models reverse malware?
Yes - sometimes extremely well. They can decode, decompile, emulate, and recover real infrastructure directly from raw binaries.
But they are not yet reliable enough to operate without expert oversight. The hardest, genuinely obfuscated samples still expose failures in strategy, verification, and adaptation - precisely where deep reverse-engineering expertise matters most.
1
2
145
We're open-sourcing MalwareBench: the tasks, human-authored ground truth, deterministic scorer, and all 252 agent trajectories.
Read full blog post: arimlabs.ai/writing/malware-…
GH Repo: github.com/arimlabs/malware-…
2
132
arimlabs.ai/writing/loss-of-…
Next: cybersecurity capability evals on the same cohort. Coming soon. Stay Tuned (7/7)
🚨 BREAKING: Autonomous browsing AI agents are no longer secure.
Over 1,500 LLM-powered projects rely on the open-source Browser Use (60k+ ⭐) framework—each one now a potential entry point for attackers.
🎥 Watch the zero-click Proof of Concept ➡️ (video attached)
Your deployed agent may already be compromised. Here's why 👇
5/
📄 Full details in our new paper: “The Hidden Dangers of Browsing AI Agents.”
Now live on arXiv. 🔗arxiv.org/abs/2505.13076
We expose how autonomous agents can be hijacked, manipulated, and used against you—with zero clicks.
Read it before your attacker does. Contact us before they do: [email protected]
6/
AI agents are browsing the internet. Attackers are browsing them back.
Unrestricted capabilities. Blind trust.
Patch your agents. Lock down your systems.
Follow @arimlabs for cutting-edge LLM exploits, red team techniques, and defense strategies.
#AIsecurity #LLM #PromptInjection #BrowserUse #Infosec #RedTeam
3
518


