Security news from the BlueTeamSec community for Friday the 23rd of October 2026. I'm Sebastian.
And I'm Tess. A busy one today — with ten stories to get through.
Amazon's threat intelligence team detected exploitation of PHP2Shell vulnerabilities in their MadPot honeypot four days before the public disclosure on the 17th, then watched mass exploitation kick off once the embargo lifted on the 21st. Worth flagging — it's another case where the window between disclosure and widespread exploitation has effectively vanished, likely aided by AI-assisted tooling.
Microsoft have written up Storm-0940, a China-aligned actor using an AI penetration testing framework called ARTEX to automate supply chain attacks at scale. The campaign started in September through a ConnectWise breach and leveraged AI for reconnaissance, password spraying, and OAuth abuse — machine-speed execution across hundreds of organisations. We covered a related defensive order from Homeland Security earlier this week targeting one of the AI models involved.
JPCERT have published an investigation into recent attacks in Japan where threat actors are reverse-engineering mobile apps, particularly Flutter-based Android applications, to extract API credentials and endpoints. They're then exploiting server-side vulnerabilities like SSRF and deploying web shells with EDR evasion. One for anyone relying on client-side security or embedding secrets in mobile apps.
Straiker have profiled a group called CyberCartel who've deployed CARLIY0S, an AI framework built on PentAGI, to attack over 400 financial institutions across Latin America. The system uses Claude-based agents in what they're calling a cognitive swarm architecture, complete with jailbreaking techniques to bypass safety controls. It's being framed as the emergence of AI-powered persistent threats.
Florian Roth has highlighted process-watcher, a Linux tool using eBPF to monitor subprocesses spawned by PHP environments — designed to detect post-exploitation activity from the PHP remote code execution attacks we've been seeing this month. It's a reactive detection mechanism rather than a patch, but adds useful visibility if you're running exposed PHP services.
Railway disclosed unauthorised access to systems storing customer environment variables, resulting in exposure of both encrypted and decrypted secrets including API keys and database credentials. The incident appears linked to a broader campaign targeting platform providers to compromise AI inference endpoints. Railway paused all deployments and are notifying affected users. This follows a similar disclosure from Zeabur we covered yesterday.
eMudhra have disclosed a certificate security incident involving eight TLS intermediate certificate authorities with weak cryptographic keys due to a procedural flaw during key generation. The weak keys were discovered by an independent researcher eighteen months after creation, leading to revocation and browser distrust of the associated root. No end-entity certificates were ever issued, fortunately.
BerriAI have confirmed that version 1.107.0 of the Python library litellm was compromised and published to PyPI using a lookalike organisation. The malicious package contained credential-stealing malware targeting cloud credentials, SSH keys, and API keys, plus a persistent backdoor that fetched remote code from GitHub. It was live for under two hours but also infected Docker images tagged as latest and main-stable.
And finally, researchers at FoxIO have disclosed Header Accordions vulnerabilities affecting HTTP/3 server implementations through QPACK decompression flaws. Unauthenticated attackers can exploit improper resource limits on decompressed headers to cause denial-of-service via memory exhaustion or CPU overload. Seven out of ten tested implementations were found vulnerable.
That's everything for today from BlueTeamSec. The writing is by the original authors — the analysis was automated.