The last few days have offered a vivid, uncomfortable preview of where AI-driven software is headed: it's becoming powerful enough to both find the internet's gravest security flaws and arm the people exploiting them, often in the very same week.

PHP's 22-year-old bug, found in days

The biggest story is PHP2Shell, a pair of critical remote-code-execution flaws discovered using frontier AI models and now triggering emergency patches across roughly two-thirds of the web's backend infrastructure. The security firm Calif says it found the first bug using OpenAI's GPT-5.6 Sol as part of OpenAI's Patch the Planet program, which pays for AI-assisted security reviews of open-source code. Separately, Wiz says it found the flaws independently using its own tooling plus Google's Gemini 4 Argon and GPT-5.6 Sol, reporting them to PHP on August 24. Wiz also noted that building the exploit took one researcher about a week with these tools — work that would previously have taken a top team months.

What makes this notable isn't just the discovery — it's the compressed timeline afterward. Exploitation attempts began on October 17, four days before PHP released fixes, and by Friday a second, related unpatched vulnerability had already surfaced. The episode is becoming a textbook case for how AI collapses the gap between "theoretically exploitable" and "actively weaponized."

The offense side of the ledger

That same dynamic is playing out on the attacker's side. Microsoft disclosed that a China-state aligned actor it tracks as Storm-0940 used the agentic AI penetration-testing framework ARTEX to automate attacks against hundreds of organizations globally, with several attacks resulting in privileged cloud access in less than three hours. The campaign, tied to a supply-chain compromise at IT vendor ConnectWise, reportedly peaked on October 11 at 2,900 failed sign-in attempts per minute, generating 111,000 failures in a single day across more than 2,250 tenants. Researchers later found the attacker framework configured with DeepSeek, GLM, and Grok via OpenAI-compatible gateways — a reminder that open, swappable AI backends make offensive tooling resilient to any single vendor's safeguards.

Anthropic bets on proactive agents

Amid the security chaos, Anthropic pushed forward on product. The company launched Claude Duet, a proactive agent that "reaches out before you ask," connecting to a user's apps and suggesting things that need attention. Built on Opus 5.5, Duet sends a short brief each morning on important items from connected apps, then checks in every few hours, while requiring explicit approval before sending a message, making a payment, signing a document, or deleting a file. It positions Anthropic alongside Meta's Muse and OpenAI's Dots in a fast-forming race to own the "always-on assistant" relationship with consumers, not just developers.

Taken together, the week's news underscores a theme that's becoming impossible to ignore: AI agents are no longer confined to chat windows. They're patching the internet, breaching it, and now quietly managing people's inboxes — sometimes all at once.

Sources

AI AgentsAnthropicArtificial IntelligenceCybersecurityDaily AI UpdatePhp2shell