/Hive Five #259 · Jan 30, 2026
The Browser is the Sandbox
/The Bee's Knees 🐝
The headline picks from this issue. The full issue has more.
-
BodySnatcher: Critical Agentic AI Vulnerability in ServiceNow
Researchers discovered CVE-2025-12420, a broken authentication flaw allowing unauthenticated attackers to impersonate any ServiceNow user using just an email address, then hijack AI agents to create admin backdoors.
Read : BodySnatcher: Critical Agentic AI Vulnerability in ServiceNow -
Copilot or Coconspirator: Tricking GitHub Copilot
Security researcher chains TOCTOU, prompt injection, and poisoned pipeline execution to steal secrets from repos using Copilot Agent. The attack requires only creating a legitimate-looking issue that maintainers assign to Copilot.
Read : Copilot or Coconspirator: Tricking GitHub Copilot -
The Browser is the Sandbox for AI Agents
Paul Kinlan explores how the browser's 30-year security model for running hostile code might be better suited for agentic AI than custom VMs, examining file system isolation, CSP network controls, and WebAssembly sandboxing.
Read : The Browser is the Sandbox for AI Agents -
Scaling Long-Running Autonomous Coding Agents
Cursor ran hundreds of agents for weeks, writing over a million lines of code to build a web browser from scratch. Key insight: separating planners from workers and using model-specific roles dramatically improved coordination.
Read : Scaling Long-Running Autonomous Coding Agents -
AI Models Now Exploit Real-World Vulnerabilities Without Custom Tools
Anthropic reports Claude Sonnet 4.5 can autonomously replicate the Equifax breach using only standard Kali Linux tools, highlighting the urgency of patching known CVEs before AI agents exploit them at scale.
Read : AI Models Now Exploit Real-World Vulnerabilities Without Custom Tools