Research & Writing
Ideas from the
safety frontier.
Technical research, threat analysis, and field notes from our work building foundational AI safety tooling.
Featuredanalysis
Feb 17, 2026
When Guardrails Fail: What Claude Opus 4.6 Reveals About Prompt Injection Risk
Anthropic's Claude Opus 4.6 system card finally quantifies prompt injection risk at scale. These numbers should reshape how enterprises deploy AI agents.
PV
Parin Vachhani
research
The Hidden Supply Chain Threat Hiding in Your AI Agent's Markdown Files
Agent behavioral configuration lives in markdown files that lack the governance of code. This creates a new supply chain attack surface.
research
How MCP Servers Turn AI Integrations Into Systemic Security Risks
The Model Context Protocol enables AI integration but carries fundamental security flaws. 43% of implementations have critical vulnerabilities.
research
The Moltbot Rush: When Viral AI Agents Expose Your Entire Digital Life
Moltbot gained 85,000 GitHub stars by promising to automate your digital life. Security researchers found it introduces risks most users don't understand.
research
Hidden in Plain Language: How Calendar Invites Became Data Extraction Tools Through Prompt Injection
A calendar event with crafted instructions could silently extract your private meeting data when you ask Gemini about your schedule. This reveals fundamental gaps in how AI systems handle untrusted inputs.
research
When AI Agents Have Privileged Access: The BodySnatcher Vulnerability Exposes a Critical Design Flaw
The BodySnatcher vulnerability shows how authentication gaps in AI agent platforms can become critical security breaches. Nearly half of Fortune 100 companies use affected systems.
analysis
When AI Democratization Meets Vulnerability: The Real Cost of No-Code AI Agents
No-code AI platforms promise accessibility. Recent research shows they also introduce security challenges traditional approaches don't address.
report
The Shadow AI Crisis: Why 40% of Organizations Will Face Security Incidents by 2030
Gartner predicts that 40% of organizations will suffer security incidents from unauthorized AI usage by 2030. Most are unprepared.
research
Cursor's Browser Just Became a Target: What MCP Server Hijacking Means for Your Security Posture
Malicious MCP servers can take over Cursor's browser, harvest credentials, and run persistent code. Learn how to protect your development environment.
research
How SuperAlign Helps Enterprises Counter AI-Powered Threats
Traditional tools cannot defend against AI-orchestrated attacks. Learn how SuperAlign helps enterprises address the critical security gaps that GTG-1002 exposed.