Top AI News - April 18, 2026

OpenAI released GPT-5.4-Cyber for defensive security work, offering broader access to verified users than Anthropic's competing Mythos model. The system can reverse-engineer software to detect malware and vulnerabilities, representing a more open approach to deploying advanced AI cybersecurity capabilities.

Google's Auto-Diagnose tool uses LLMs to automatically diagnose integration test failures with 90% accuracy, successfully deploying across 52,635 tests and ranking among the top 4% of internal developer tools. The deployment validates that LLMs can effectively reduce the cognitive load of debugging complex systems when integrated into existing developer workflows.

Hugging Face

VAKRA is a new executable benchmark testing AI agents on complex, multi-step enterprise workflows across 8,000+ APIs and 62 domains. Leading models struggle significantly with compositional reasoning, multi-hop tasks, and policy constraints, exposing major gaps in agent reliability for production use.

TechCrunch

Trump administration officials are encouraging major U.S. banks to test Anthropic's Mythos model for vulnerability detection, even as the company battles the DoD in court over supply-chain risk designations. The mixed government approach highlights conflicting regulatory stances on AI deployment in critical infrastructure.

arXiv

A security study of 428 third-party LLM API routers found widespread malicious activity including code injection, credential theft, and cryptocurrency drainage, exposing critical vulnerabilities in the AI agent supply chain that currently lacks cryptographic integrity protections.