Two critical PHP vulnerabilities and a separate AI-assisted cybercrime campaign targeting financial institutions have raised security concerns in October, as attackers exploit software weaknesses and use artificial intelligence to scale their operations.
The PHP vulnerabilities, tracked as CVE-2026-109618 and CVE-2026-109619 and collectively known as PHP2Shell, can allow unauthenticated attackers to execute arbitrary code on vulnerable servers. The PHP project released security updates on October 21, 2026, after details of the flaws became public ahead of the planned disclosure.
The vulnerabilities affect PHP’s request-processing and memory-management mechanisms. CVE-2026-109618 involves memory corruption during out-of-memory handling, while CVE-2026-109619 involves stale internal state referencing memory that has already been freed. Under vulnerable configurations, attackers can exploit these conditions to run code with the privileges of the web server.
The PHP project considers Apache deployments using mod_php particularly exposed, although other web-serving configurations, including PHP-FPM, php-cgi, FrankenPHP, and LiteSpeed, can also be affected. Since PHP powers a large number of websites, web applications, and embedded systems, administrators need to check the actual PHP runtime rather than relying only on whether their content management system or plugins are updated.
The security releases include PHP 8.5.12, 8.4.27, 8.3.36, and 8.2.35. Administrators running older, unsupported versions should check for vendor-provided security backports or plan an upgrade. Organizations that cannot immediately apply a fix should follow vendor guidance on restricting access to exposed services and applying temporary mitigations.
Security agencies and technology vendors have also reported exploitation attempts and urged organizations to prioritize remediation. Some organizations have taken public-facing services offline temporarily while assessing their exposure and applying updates.
Separately, cybersecurity company Straiker reported that a cybercrime group known as CyberCartel used an AI-assisted system called CARLIY0S to target more than 400 financial-sector organizations in Latin America. The targets reportedly included banks, payment providers, and fintech companies.
According to Straiker’s investigation, the operation used multiple AI models, including Anthropic’s Claude, xAI’s Grok, and models from China’s Moonshot AI. The system assigned different tasks to AI agents, including assessing targets, developing attack plans, and collecting credentials and other sensitive information.
The researchers reported that the operators organized AI agents into a coordinated workflow, with some models handling attack planning while others gathered information and collected exposed secrets. The group also reportedly attempted to get around model safety restrictions by presenting malicious tasks as authorized penetration testing, academic exercises, or vulnerability research.
A snapshot of a database called FinancialSoft, examined on September 24, 2026, contained 4,493 records, according to Straiker. The inventory included 2,911 email addresses, 903 bank account numbers, and 391 API keys, alongside passwords, login credentials, database connection strings, cryptographic keys, and other sensitive information.
The researchers also reported finding more than 270 penetration-testing reports and information about over 200 vulnerabilities associated with the group’s activities. These materials were separate from the 4,493-record database snapshot.
The reported target count does not establish that every organization was successfully compromised. Straiker’s findings also do not mean that every credential in the inventory remained valid or that all affected accounts were accessed. The financial institutions were not publicly identified in the reporting.
Straiker said it shared information about the activity with the AI providers. The company reported that Anthropic revoked the credential used to access Claude, while xAI began investigating the activity associated with Grok. The findings primarily reflect Straiker’s investigation, and the extent of the confirmed compromises remains unclear.
For organizations responding to the PHP vulnerabilities, the immediate technical priority is to identify affected PHP installations, install the appropriate security updates, restart relevant workers, and review logs for suspicious requests, unexpected child processes, newly created PHP files, and unusual outbound connections.
If this article helped you, please consider supporting our work. Every small contribution keeps Abijita.com independent and running.
Security teams investigating possible AI-assisted intrusions should also review privileged account activity, API key usage, authentication logs, and unexpected access to financial systems. Credentials found in exposed or compromised environments should be revoked or rotated, and affected systems should be investigated for unauthorized changes.
