Updated 3 minutes ago

Share this article

Related News

ChatGPT Ads expands to 26 more countries, reaching 89 markets

Oct 21, 2026

ChatGPT Ads expands to 26 more countries, reaching 89 markets

OpenAI says more than 100,000 advertisers have used ChatGPT Ads and is extending the product to 26 additional countries. Ads remain limited to Free and Go plans, while the company says conversations are private from advertisers and placements stay separate from answers—claims users should assess through the controls and disclosures they actually see.

ChatGPT AdsOpenAIadvertising

OpenAI models are coming to Google Cloud through Model Garden

Oct 21, 2026

OpenAI models are coming to Google Cloud through Model Garden

OpenAI and Google Cloud plan to offer frontier models and Codex workflows through Model Garden, with Google Cloud Marketplace billing and support for existing GCP spend commitments. General availability is still weeks away, so teams should separate announced distribution from live regional and model-level availability.

OpenAIGoogle CloudModel Garden

OpenAI buys Orchestra Bio to add workflow agents to its drug-development stack

Oct 20, 2026

OpenAI buys Orchestra Bio to add workflow agents to its drug-development stack

OpenAI has acquired Orchestra Bio and plans to integrate its drug-program coordination software with Rosalind, GPT-Rosalind and Rosalind Workbench. The deal expands OpenAI from research models into the operational layer around medicine development, but the price, product continuity and any effect on treatment timelines remain undisclosed or unproven.

OpenAIOrchestra BioRosalind

PHP2Shell patches arrive as AI-assisted discoveries trigger a web security response — Source: OpenTools Team / original PHP2Shell: the runtime needs a fix evidence mapSource image

AI and Web Security

PHP2Shell patches arrive as AI-assisted discoveries trigger a web security response

By OpenTools Team

PHP has released fixes for two critical flaws that can let attackers take over vulnerable web servers. Researchers at Calif, OpenAI and Wiz helped establish the danger using AI‑assisted work, while website owners now need to verify their hosting runtime and distinguish attempted attacks from confirmed breaches.

The fix is in PHP, beneath the website

The PHP project released security updates on October 21 for CVE‑2026‑109618 and CVE‑2026‑109619, two flaws rated 9.2 out of 10. Under exploitable conditions, someone sending web requests without logging in can run code on the server. The problem sits in PHP, the software that executes many websites, rather than a particular theme or plugin.

One flaw concerns cleanup after a request runs out of memory. The other leaves references to memory that a later request may use after it has been freed. PHP flags Apache's mod_php as the highest‑risk configuration, with exploitation demonstrated before website code runs. PHP‑FPM and other web‑serving installations also need assessment.

The fixed supported releases are 8.5.12, 8.4.27, 8.3.36 and 8.2.35. A distribution's backported package can contain the fix without adopting those exact upstream numbers, so the vendor advisory matters. The practical question for a website owner is which runtime is actually serving visitors.

AI helped establish the danger; maintainers delivered the fix

The PHP Foundation's account credits Calif researcher Ryan A, OpenAI Codex Security including Amy Burnett, and Wiz with research and reporting. PHP maintainers led the patch work. The Foundation says earlier signs of the memory problem emerged through work associated with a 2024 Quarkslab audit organized through OSTIF, but difficult compatibility choices and limited specialist capacity delayed a complete fix.

Wiz describes using Atlas, Gemini 4 Argon and GPT‑5.6 Sol to investigate and adapt exploits across test applications. That is a concrete account of AI‑assisted security work, while claims about weeks or months saved remain the researchers' own assessment. It does not establish that an autonomous model independently managed the response.

The consequential AI development is that weaknesses regarded as difficult to turn into attacks have now been demonstrated in realistic deployments. The same finding creates work for defenders: validate the report, develop a compatible patch, distribute it and prove that operators have installed it.

Attack reports require a careful distinction

Cloudflare says it observed attack traffic and deployed default protections for customers with its web application firewall enabled, including free‑plan customers. It also says attackers changed their requests to try to evade protections. These are provider observations, not a public census of compromised websites. The sources reviewed for this report do not identify a responsible attacker group or establish a reliable successful‑breach total.

The PHP Foundation says leaked identifiers and a mistakenly published demonstration video disrupted coordinated disclosure, prompting a release one day earlier than planned. Its account also describes an early distributor release that let others inspect the fix before the coordinated publication. Those events help explain the urgency; they do not establish who initiated attacks.

A precautionary shutdown is similarly not proof of a breach. For customers and employees, the relevant follow‑up is a service operator's verified statement about affected data and restoration, rather than treating every outage as confirmed theft.

What website owners should ask their host now

The UK NCSC advises organizations to inventory PHP, prioritizing internet‑facing systems and products that bundle it. A business using managed hosting should ask for the affected runtime, the vendor fix applied, the deployment time and confirmation that the running service was restarted or replaced. WordPress administration alone may not expose those answers.

If a prompt update is impossible, follow the current PHP and hosting‑provider mitigation instructions. NCSC says taking a system offline should be considered when temporary protections cannot be applied adequately or quickly. Configuration changes and a firewall can reduce exposure to particular paths, but PHP says an update is the complete solution.

For potentially exposed systems, preserve logs and assess for compromise as well as patching. If an investigation finds unauthorized access, involve incident responders and revoke affected secrets from a clean environment. Patching closes the entry point; it does not by itself remove a backdoor already installed.

The next test is deployment and accountable follow‑through

The immediate story now moves from disclosure to deployment: which hosts and appliances ship updates, whether their customers receive them, and whether investigators confirm real compromises. Useful next evidence would include exact affected versions and artifacts, documented attack outcomes and provider‑specific recovery instructions.

For the AI industry, this also tests the capacity behind coordinated disclosure. Faster discovery increases the value of experienced maintainers and reliable release channels. It does not eliminate the time needed to choose a safe fix or verify it across the software ecosystem.

Figure by OpenTools Team; no third‑party expressive material used. Original explanatory figure, based on the linked PHP and research advisories. It separates discovery, the runtime fix and the operator response; it is not a photograph of an attack.

Related: How agent access can reach live personal and confidential data.

Sources

1.UK NCSC advises organizations to inventory PHP (ncsc.gov.uk)
2.php.net (php.net)
3.thephp.foundation (thephp.foundation)
4.blog.cloudflare.com (blog.cloudflare.com)
5.wiz.io (wiz.io)

Tags

PHP2ShellAI security OpenAIWizPHP