It’s happening. The AI-exploit era just arrived at your front door.
Three months ago, our partners Softwire wrote that Mythos-class AI capability would turn attack timelines from months into hours. This month, it stopped being a prediction and exploits discovery by AI made worldwide headlines.
On October 21st, Cloudflare recorded tens of thousands of exploit attempts within 45 minutes of PHP2Shell (CVE-2026-109618) going public, hundreds of thousands within two hours, and millions by that evening. PHP2Shell is a zero day in the PHP runtime, more on what that means below. The flaw sits in PHP builds over the last decade, back to PHP 5.4, and touches an estimated 70% of the web. It was discovered by AI, not humans, as part of OpenAI’s Patch the Planet initiative.
The same day, Node.js shipped emergency fixes for Node2Shell (CVE-2026-111497), another zero day exploit, this time a critical remote code execution flaw with a CVSS score of 9.8. Also found by AI, specifically OpenAI’s Codex Security tool. The flaw turns out to affect Python and a long list of other libraries.
What’s a zero day, and why does this one matter?
A zero day is a software flaw the vendor doesn’t know about yet. Attackers who find one get a free run until a patch exists. For years, that window ran in weeks or months, long enough for responsible disclosure, patch, test, deploy. The PHP exploit closed that window. The bug sat in the PHP source for more than 20 years. AI found it in weeks of focused work & Other AI weaponised it within hours of public disclosure. Meaning the bar for someone to exploit vulnerability became an API key for your favorite AI tools.
The middle of the market has nowhere to hide
Big banks fund defence in depth. The largest banks and government organisations sit inside Project Glasswing and get to patch ahead of disclosure. Councils, scaleups and Fintechs do not. PHP2Shell has the potential to land on any of them. Node2Shell has the potential to land on any of them. The next five exploits after that will too.
Fintechs sit in the sharpest position. Lower cost tech means lower resilience. Lower resilience means your product is where the AI probes land first. Consumer trust in a post-Mythos market will favour whoever is visibly secure. The firms that come out of the next twelve months with reputation intact will take share from firms that don’t.
The CSA top three are now imperative
So what can you do about it, when it feels inevitable that AI agents working for weeks or months on exploits are coming either way. In February The joint CSA, SANS and OWASP briefing lays out three actions that matter above all others.
1. Point AI agents at your code. LLM-driven security review on every change before merge. Agentic scanning embedded in CI/CD.
2. Adopt AI agents in your secure coding practices. Tools, prompts, evaluations and guardrails inside your engineering workflow, not a policy document on a shared drive.
3. Defend your agents. Audit prompts, tool definitions, retrieval pipelines and escalation logic. Set scope, blast radius and override before anything goes live.
Each is engineering work in your repos, your pipelines and your agent stack. None of them demand a seven-figure security transformation or a pen test. All three are achievable inside your existing engineering team with the right skills and support.
Where Softwire helps
Softwire has a 25 year history of enabling companies to get the most of technology safely, and supports financial services institutions and government departments in exactly this. Email FinancialServices@Softwire.com to open a conversation on how we can support you with the CSA top 3 on your product developments.
