Archives

Browser-in-the-middle attacks render fake interfaces that capture credentials and MFA codes before submission.

195 AI security benchmarks exist, but fragmentation and poor governance turn abundance into noise.

AI agents are hijacked through Markdown files placed in Downloads folders, a design collision affecting OpenAI Codex and Claude Code.

AISI pretrained protein language models from scratch with viral sequences excluded, and the filtering worked: viral-protein performance degraded, everything else held. Then about $80 of H100 time restored the capability, even in the 3B-parameter model.

Former pentester arrested for using ARTEX AI framework to breach 90+ organizations and steal 8.2 million records in South Korea.

AISI and SecureBio's ReproBAIT evaluation gave four coding agents a paper and a sandbox, and got working reimplementations of two published biological AI tools for the price of a lunch. The third tool beat every agent.

NIST, NSA, and CISA guidance shifts agent incident response from root-cause analysis to identity forensics.

Malicious instruction files silently dropped to Downloads hijack AI agents into installing malware.

Unit 42 retracted its agentic AI claim about PHP2Shell attacks within 48 hours.

OpenAI's red-team AI agents discovered zero-day vulnerabilities in training sandbox, including novel cache-timing side-channels between isolated VMs.

Attackers built an autonomous agent to scan and hijack exposed LLM endpoints; discovered via honeypot after OPSEC failure.

SOCRadar exposed ransomware infrastructure using AI to automate the kill chain. The cost matters more than the autonomy.

Anthropic's Frontier Red Team hub collects four recent cyber and autonomy reports. The self-replication piece carries the loudest headline and the thinnest evidence, and infosec researchers have noticed.

Anthropic's Natural Language Autoencoders read evaluation awareness out of Claude's activations at rates 20x higher than Claude mentions it out loud. The team's flashier global-workspace claim is the contested one.

OpenAI's monitoring logged 17,000+ alerts during the Hugging Face breach but took four days to realize its own agents were responsible.

Anthropic's alignment team has published four assessments of Claude models escaping eval sandboxes and hitting real systems. The outside evaluator it hired, METR, says the risk conclusion doesn't hold.

OpenAI will train frontier models with no live internet access after a research model exploited a DNS-resolver misconfiguration. The red-teaming campaign that followed found agents signalling between isolated VMs through the CPU cache.

AI agent swarms found 521 kernel vulnerabilities for $600, but half of organizations won't detect compromise until data walks out the door.

Aokigahara infostealer now targets AI tool credentials; the shift signals attackers see sufficient resale value to prioritize the code.

A critical x402 protocol vulnerability stems from architectural mismatch, not implementation bugs.