PHP Vulnerabilities Exposed by AI Models, OpenAI Expands Cyber Defense Business

iconFLASH AI
Share
AI summary iconSummary
FLASH AI reported in August 2026 that AI models uncovered two critical PHP vulnerabilities, php2shell, affecting 70% of the web. Patches were released October 20, but attacks started October 17. OpenAI’s GPT-5.6 Sol and Google models aided the discovery. OpenAI is expanding its Daybreak program and plans a cybersecurity-focused model this fall. On-chain data shows rising interest in altcoins to watch amid heightened security concerns.

This roundup covers PHP security vulnerabilities, OpenAI's cyber defense business and other AI developments. Multiple research teams at different organizations used AI models in August to identify two vulnerabilities in PHP, which supports an estimated 70% of the web. The vulnerabilities and the technique for exploiting the vulnerabilities are collectively called php2shell. OpenAI's GPT-5.6 Sol appears to have contributed significantly to the discovery, while other models, including Google's Gemini 4 Argon, also appear to have contributed.

The PHP Foundation announced on October 20 that a critical patch would arrive within 24 hours. A coalition including OpenAI, Wiz and Calif publicly disclosed the vulnerabilities the next day after privately collaborating on a fix. Information about the vulnerabilities and exploitation techniques appears to have leaked before the foundation's announcement. Amazon Web Services said hackers began exploiting php2shell on October 17, before public disclosure. Organizations notified in advance, including Amazon, had agreed not to publish information before 14:00 UTC on October 21.

The interval between the patch announcement and release left a day-long exposure window, compounded by delays as companies updated systems. Millions of attempts to exploit the vulnerabilities followed public disclosure, and affected websites continue installing the patch. WordPress, the service most affected by the vulnerabilities, reported no problems. Many governments and some private entities temporarily disabled web services while awaiting patch installation. The shutdowns prevented damage to those sites but also stopped people from using the services.

More than 20 cybersecurity incidents occurred in October, with many happening this week. Emily Forlini argues that organizations must use AI to identify and repair vulnerabilities before attackers exploit the vulnerabilities. Experts said this week that warnings about increasingly capable AI helping attackers exploit software bugs are becoming reality. AI models have improved at processing large codebases and acting independently, while coordinated groups of AI agents can operate faster than humans can follow.

OpenAI has been expanding Daybreak, an application-only program giving organizations access to advanced tools for defending against sophisticated AI-powered attacks. OpenAI updated Daybreak's website two weeks ago, raising the reported number of flaws discovered by participants from 858 to 1,885. The reported number of patches increased from 263 to 661. OpenAI co-founder and president Greg Brockman promoted the program through a post about the PHP exploit and OpenAI's contribution to discovering the vulnerabilities. Brockman also hired a chief revenue officer with cybersecurity experience over the summer and assigned the executive to build a sales team for defensive software.

OpenAI plans to introduce a cybersecurity-focused model this fall. OpenAI also plans a first-of-its-kind tool intended to simplify deployment of the model across businesses' infrastructure. Forlini argues that AI-powered cyber defense is becoming a revenue opportunity, provided attackers do not compromise the providers first. Anthropic offers Project Glasswing to cybersecurity professionals at vetted organizations, and other software-as-a-service providers also sell AI-powered cyber defense tools.

Three people familiar with Anthropic's plans told Yahoo Finance that Anthropic could publish the prospectus for its planned IPO as soon as tomorrow. The offering could be the largest ever and could value Anthropic at more than $2 trillion.

AI safety firms Redwood Research and METR found that 10% of internal traffic to OpenAI's Codex coding tool escaped monitoring for months this spring. Redwood Research and METR worked inside OpenAI for an independent security audit over three weeks, from July 21 to August 11. The firms found a monitoring gap from May to July that agents could have exploited to behave improperly without detection, alongside numerous other monitoring weaknesses. The period overlapped with the Hugging Face cyber incident, but the audit did not link the monitoring failure to that incident.

The roundup also covers Anthropic introducing a preview of Duet, a personal AI agent competing with Instinct, Meta's Muse and OpenAI's Dots; Microsoft reporting hundreds of autonomous AI-powered hacks by China state-aligned group Storm-0940 using the China-developed open-source platform Artex, also used in a hack of South Korean banks earlier this month; and Barclays planning 5,000 job cuts, equivalent to 5% of the workforce, in an AI-driven cost-reduction effort after quarterly profit rose 12% but missed analyst forecasts.

Disclaimer: The information on this page may have been obtained from third parties and does not necessarily reflect the views or opinions of KuCoin. This content is provided for general informational purposes only, without any representation or warranty of any kind, nor shall it be construed as financial or investment advice. KuCoin shall not be liable for any errors or omissions, or for any outcomes resulting from the use of this information. Investments in digital assets can be risky. Please carefully evaluate the risks of a product and your risk tolerance based on your own financial circumstances. For more information, please refer to our Terms of Use and Risk Disclosure.