Forbes reported on October 23, 2026, that OpenAI’s GPT-5.6 Sol helped researchers identify serious PHP vulnerabilities through Patch the Planet, an OpenAI-led security effort. The PHP project’s release notice names PHP 8.5.12 as the fix for two vulnerabilities; AWS separately reported an attempted exploit before public disclosure, without establishing success on real systems.
How OpenAI’s work connected to the PHP flaws
Forbes reported that researchers used GPT-5.6 Sol through Patch the Planet, an initiative involving OpenAI’s internal and external security researchers. OpenAI and its collaborators spent more than two months validating findings and fixes before disclosure, according to Forbes.
The PHP development team announced PHP 8.5.12 on October 21, 2026. Its security release notice says the update fixes CVE-2026-109618 and CVE-2026-109619, vulnerabilities that can allow unauthenticated remote code execution. The project called upgrading the only safe choice for PHP 8.5 users. NeoTeo’s earlier coverage of the PHP fixes provides the release details; this article adds the context around OpenAI’s program and AWS’s observation.
What Daybreak Blue and Red allow
Project Daybreak, also called Trusted Access for Cyber, gives approved users access under two tiers. OpenAI’s published program rules describe different model access and approval requirements:
| Access tier | Model access | Approval and conditions |
| Blue | Fewer refusals on supported general-purpose models for verified defensive work. Blue does not include cyber-specialized models; standard protections remain for Astra. | Use is limited to verified defensive work. |
| Red | Access to selected cyber models and fewer refusals on certain supported models. | Requires separate approval and stricter verification; some models require additional approval. |
OpenAI limits Daybreak work to systems users own or operate, or have explicit permission to test or analyze. The program does not allow work for external customers or third-party workflows. Blue therefore means narrower restrictions for specified defensive work, not unrestricted cybersecurity access.
What AWS observed before disclosure
AWS said its MadPot telemetry recorded a PHP2Shell exploitation attempt on October 17, four calendar days before public disclosure on October 21. AWS’s account of the activity says it is unclear how the actor obtained exploit information and whether the observed attempts would have succeeded on real systems.
Why finding flaws does not finish the repair work
Forbes reported that 1,316 of 1,995 Patch the Planet findings were still awaiting patches in the snapshot it described. It also reported on an audit of 130 deployed AI-written patches: 16 had functional problems, and one had a performance issue. The audit figures and backlog were reported by Forbes; they point to the work that remains after discovery, including validation, correction, and deployment.
