Blog
Research, perspectives, and field notes on where security is heading.
The Loop
AI-assisted development creates security triage bottlenecks that faster scanners cannot solve; routing findings to service owners with drafted fixes is the actual scaling lever.
· Cantina
The Loop
AI-generated code often contains authorization flaws that static analysis misses because the vulnerabilities exist across multiple functions rather than within individual code blocks.
· Cantina
The Loop
A poisoned VS Code extension compromised a GitHub employee device, exposing 3,800 internal repos in a broader supply chain attack campaign.
· Cantina
The Loop
Governance requires mapping policy clauses to runtime checks and audit evidence, not just writing policies.
· Cantina
The Loop
Analysis of Apple's May 2026 security patch cycle that closed 80+ vulnerabilities, including AI-credited findings from both Cantina and Anthropic.
· Cantina
The Loop
Four production-critical AI infrastructure vulnerabilities reveal systemic gaps in how organizations manage interconnected security risks across their AI stack.
· Cantina
The Loop
Analysis of Dirty Frag, a Linux privilege escalation vulnerability affecting major distributions through networking subsystem flaws.
· Cantina
The Loop
Cyber-enabled cargo theft reached $725M in losses during 2025, exploiting compromised dispatch workflows and freight software to orchestrate sophisticated theft schemes.
· Cantina
The Loop
Cantina's autonomous agent Apex discovered a memory safety vulnerability that persisted in WebKit for 13 years, along with two CSP bypasses, all patched in iOS 26.5.
· Cantina
The Loop
Governance of AI agents is shifting from policy frameworks to real-time runtime controls as regulatory bodies establish standards and minimum security requirements.
· Cantina
The Loop
Federal agencies warn of active exploitation of internet-connected PLC controllers across U.S. critical infrastructure, creating operational trust and safety concerns.
· Cantina
The Loop
Prompt injection becomes a critical security issue when AI agent frameworks allow model-controlled input to reach file operations or code execution.
· Cantina
Page 2 of 25 · 294 posts