About

A malicious repository used its own Claude Code subagent configuration to send the first-pass security review to the smaller Haiku model, with a scope that skipped the tests directory. Asked to prepare a pull request, the agent followed the repo’s documented workflow and ran pytest, executing a payload hidden in a test file that launched a Sliver implant while reporting only a passing test run. No jailbreak or prompt injection needed: the attack abuses the trust placed in repo-defined agents, workflow instructions and test execution.

• Claude Code
• AI Coding Agents
• Supply Chain
• Sliver C2

Straiker STAR Labs research covered by Forbes; I was part of the research team behind it. A polite email with step-by-step “organize your Drive” instructions gets followed by Perplexity’s Comet AI browser when the user asks it to complete their recent email tasks, moving critical files to trash at scale through its Gmail and Google Drive access. No jailbreak or traditional prompt injection: the attack succeeds by sounding like routine productivity work.

• AI Browsers
• Agentic Abuse
• Social Engineering
• Google Drive

Proposes a machine-learning classifier to detect SQL injection attacks and makes it interpretable through counterfactual explanations (XAI), surfacing the minimal changes to a query that would flip the model’s decision. It brings explainable AI to database security so analysts can understand and trust alerts instead of treating them as a black box.

• XAI
• Counterfactuals
• SQL Injection
• Machine Learning

Applies computer vision with YOLOv8 to automatically identify Triatoma infestans, the main vector of Chagas disease, training on 91 labeled images and 9,100 augmented ones. It reaches a mAP@50 of 0.9588 and tells the vector apart from similar insects with no false positives, laying the groundwork for entomological surveillance on mobile and embedded systems.

• YOLOv8
• Computer Vision
• Chagas
• Vector Surveillance