Louis Columbus

Cisco, Box, and Intuit security leaders at VB Transform 2026 detailed how multi-turn attacks expose gaps that single-turn AI red-teaming still misses.

A new ransomware strain scans specifically for AI model weights, and it has no way to collect a ransom, leaving victims facing a $500,000 rebuild.

Rajat Taneja explains why Visa open-sourced the AI harness that found critical bugs in its payment network, and the new metric it built to verify fixes work.

AI model verification has relied on ClamAV scan badges and a repository tag no one has to prove. Cisco's new tool grounds lineage in weight-level fingerprinting instead.

For decades Mastercard's fraud rules were built to stop bots. Now AI shopping agents are buying, forcing Mastercard to rewrite its core risk framework.

A hijacked GitHub account let the Shai-Hulud worm pass npm's trust check, spreading through packages with 2 billion monthly installs and stealing CI secrets.

Four of five enterprises that secured AI agent identities never built isolation to contain a compromised agent, VentureBeat's July Pulse survey found.

Anthropic gave three Claude agents conflicting orders on one server. They sabotaged each other, disguised malware, and hid it from the humans relying on them.

An AI agent read a blocked Cloudflare log as an order, rewriting a company's DNS nine times out of ten. The fix: propose changes, don't approve them.

Visa's updated security harness now writes and merges its own code fixes by default, with no human review required before patches reach production.