IA · 7 November 2026 · 5 min read

Autonomous Hacking Goes Budget: Open Source Closes In and Exposed Agent Tools Proliferate

In brief: A series of independent evaluations by AI benchmark firm Vals reveals that Chinese open-weight models such as DeepSeek and GLM have virtually erased the offensive gap with locked frontier systems. Meanwhile, internet scans conducted by Censys uncovered more than 800 deployments of the autonomous agent tool Artex left wide open without password protection, handing formidable automated penetration testing resources to anyone browsing the web.

by Team Mocchi's

Autonomous Hacking Goes Budget: Open Source Closes In and Exposed Agent Tools Proliferate

The window of advantage that top frontier laboratories hoped to maintain over malicious automated capabilities has effectively closed. For months, major labs including OpenAI and Anthropic kept their dedicated cybersecurity models restricted to trusted partners, operating under the assumption that gated access would afford defenders crucial time to fortify digital infrastructure before open weights caught up. Fresh industry data shows that buffer has now evaporated, while the unmonitored deployment of automated offensive tooling is already generating friction across the public web.

Open weights vs. proprietary labs: closing the capability gap

The most rigorous assessment of this shift comes from AI evaluation firm Vals, which designed a battery of network intrusion tests across simulated corporate environments. As reported by WIRED, researchers tasked various language models with multi-stage compromise workflows: navigating initial web application perimeters and executing lateral movement across isolated internal network segments.

The findings point to an increasingly level playing field. OpenAI’s proprietary cyber model, Daybreak, which remains restricted to vetted partners, successfully fulfilled 81% of assigned objectives. Yet China's open-weight model DeepSeek trailed closely behind at 73%, matching the proprietary giant on multiple individual challenge tasks.

Crucially, economics heavily favor the open landscape. Running an automated intrusion routine through leading commercial APIs can tally between $17 and $46 in token expenses per attempt. Replicating the equivalent workflow via open-weight architectures like DeepSeek costs merely $2 to $3 in compute, an expenditure that drops to near zero for operators with self-hosted hardware.

The Artex fallout: more than 800 exposed attack nodes

While model capabilities present structural challenges, the operational mishandling of agent frameworks introduces immediate risk. According to reporting from TechCrunch, threat intelligence researchers at Censys discovered over 800 active installations of Artex publicly accessible across the web without requiring any password or credential verification.

Artex, an open-source autonomous penetration testing suite introduced last July, was previously profiled by Microsoft researchers as an agentic system capable of executing full offensive lifecycles from initial reconnaissance through exfiltration with minimal human guidance. The Censys telemetry reveals that dozens of practitioners and independent teams deployed these potent agent frameworks onto public cloud compute without basic authentication barriers, leaving operational consoles, custom intrusion modules, and active target logs completely open to discovery.

The end of the defender's grace period

The combination of capable open-weight reasoning engines and turnkey orchestration scripts fundamentally alters the threat model for internet-connected systems. Advanced reconnaissance and vulnerability chaining no longer demand seasoned offensive engineers; they can be packaged and triggered autonomously at negligible cost.

Western research institutions originally predicated limited rollouts on buying defense teams a 12-to-24-month runway. With open models rapidly bridging the technical performance divide, maintaining closed-door dominance provides only marginal latency. Network administrators and software engineering teams can no longer assume that advanced adversary simulations remain confined to well-resourced laboratories.

Mocchi's take

For engineering teams and technology leaders across Europe, this reality invalidates the premise of episodic vulnerability scans and periodic security audits. When autonomous agents driven by open-weight models can relentlessly probe corporate edges for single-digit dollar budgets, defense mechanisms must match that persistent frequency. In software development, enforcing strict zero-trust baselines, segregating API tokens across deployment pipelines, and continuously reducing public perimeter surfaces are no longer secondary precautions—they are fundamental survival parameters against non-stop automated discovery.

Further reading

All articles on the Mocchi's blog