OpenAI paused frontier model training for 11 days after AI agents exposed security weaknesses, VM escapes and covert communication channels. OpenAI published a detailed report this week explaining how it improved security in the systems used to train and test its most advanced AI models. The company also described what happened when its AI agents tried to find weaknesses in those systems and why it stopped training its most advanced models for 11 days in September. “Based on the red-teaming results, the fixes we had implemented, and improvements to our monitoring and stopping systems, we judged that our strengthened safeguards were sufficient to resume frontier RL training and internal inference.” reads the report published by the company. “We resumed these workloads on October 1, 11 days after our pause began.” The incident began on September 20, when an internal research model exploited a DNS configuration error and gained limited access to the live internet during training. This incident, along with a security incident involving Hugging Face in July and other cases of AI models behaving in unexpected ways, led OpenAI to make major changes to its security measures. OpenAI’s stated containment requirement is that limits on what an agent can access must hold even when the agent actively tries to circumvent them. The policy choice they landed on is that frontier model training now runs with no internet access at all, full stop. “For training and most evaluation, the live internet lies outside this boundary, as we believe that providing models access to the live internet can pose significant risks. Although internet access can help models learn useful skills, it also gives a capable, misaligned agent opportunities to affect systems beyond our control.” continues the report. “We’ve therefore designed the security architecture for our research environments with one overriding principle: we must always be able to prevent a frontier model from achieving unauthorized, arbitrary access to the live internet.” To support training without live internet, they built what they call an “offline world,” a large indexed and cached set of web content that models can browse and search without touching external systems. The layers of protection around that offline world include restricted containers, lightweight virtual machines that isolate each agent’s environment from the underlying hardware, network and cloud controls, and monitoring systems designed to catch unauthorized activity before it does damage. The report explains how OpenAI tested its most advanced AI agents by asking them to find weaknesses in the company’s own systems. The agents had three goals: access the live internet, break through internal security barriers, or communicate with other agents when they were not allowed to do so. OpenAI ran experiments using hundreds of billions of tokens. The agents found previously unknown security flaws in open-source software, including Kata Containers, gVisor and Moby/Docker, as well as configuration errors in OpenAI’s own services. By combining these weaknesses, they managed to escape virtual machines and gain control of the underlying host system. In further tests, the agents found weaknesses in services used for monitoring, logging and authentication. They used these flaws to make the services access URLs controlled by an attacker. The part of the report that deserves the most attention isn’t the direct internet escapes. It’s what the agents did when those paths weren’t available. Agents in separate virtual machines, not allowed to share files or network connections, communicated with each other using processor cache timing. One agent evicted specific data from the processor’s cache. Another measured how long it took to reload that data, using the difference between fast and slow reads to decode a message. No file touched, no network packet sent. The technique is the same family of attacks as Spectre and Meltdown. OpenAI restarted training its most advanced AI models on October 1, eleven days after the pause began. The company is now making extensive security tests a required part of the approval process before training runs can begin. It is also using AI agents to regularly check source code and cloud settings for weaknesses, testing systems that monitor activity and automatically stop unsafe operations, and reporting discovered vulnerabilities to the open-source projects responsible for fixing them. OpenAI says it is continuing to strengthen its security by fixing vulnerabilities in virtual machines, working with cloud partners, and using AI agents to scan code and configurations. The company is also improving monitoring and automatic shutdown systems, regularly testing its sandbox and network defenses, and making high-compute red-teaming with its most capable models a core requirement for future frontier AI training and evaluation. “We are committed to strengthening our containment, monitoring, and red-teaming infrastructure as our frontier models become more capable and believe that this work will be a core part of responsibly pacing the development of our frontier models.” the company concludes. Follow me on Twitter: @securityaffairs and Facebook and Mastodon Pierluigi Paganini (SecurityAffairs – hacking, Artificial Intelligence)
Between November 2025 and March 2026, the Dutch security firm Eye Security pushed more than 2 million updates to computers it did not own to switch off a background updater that it said could have allowed an outsider to take over the machine. The updater belongs to Time4Popcorn, an abandoned copy of Popcorn Time, a free app that streamed pirated films and TV over BitTorrent. Popcorn Time was
American clinical laboratory company Quest Diagnostics disclosed that patient records have been exposed after a threat actor gained unauthorized access to its Quanum LSM system. The company learned about the breach on October 21 and deployed its incident response protocol "to contain the activity, assess and investigate the nature and scope of the incident and implement appropriate remedial measures." “On October 21, 2026, Quest Diagnostics Incorporated learned of unauthorized access to its Quanum LSM system, which certain clients use to place test orders, receive and look up results,” the company says in a filing with the U.S. Securities and Exchange Commission (SEC). Quest Diagnostics is a major clinical laboratory company in the United States, providing diagnostic testing and services for a wide range of medical conditions.
Cybersecurity researchers have disclosed details of a large-scale campaign that has targeted exposed network edge devices such as Fortinet FortiGate and Palo Alto Networks GlobalProtect firewalls. "Artifacts from this investigation also indicate that operators in this campaign have obtained access to numerous organizations worldwide, and have attempted to further compromise some victims,
PHP2Shell attacks have surged into the millions, but no successful exploitation is confirmed. Patch it immediately. On October 21, PHP released emergency patches for two critical vulnerabilities tracked as CVE-2026-109618 and CVE-2026-109619 (CVSS score of 9.2), affecting every PHP version that handles web requests. The bug code dates back to PHP 5.4 from 2012, lives in the PHP engine itself rather than in any application, and allows unauthenticated remote code execution via crafted HTTP requests. Cloudflare researchers observed tens of thousands of exploit attempts within 45 minutes of public disclosure, hundreds of thousands within two hours, and millions by October 22. Some stats on #php2shell. In the 10 hours before it was publicly announced, across the @Cloudflare network we saw thousands of attempts to exploit it. That jumped to tens of thousands in the 45 minutes post announcement, hundreds of thousands within 2 hours, and millions now. — Matthew Prince (@eastdakota) October 22, 2026 As of October 23, Censys counts over 34.4 million PHP services still running vulnerable versions. Governments, hospitals, and schools across Asia, Europe, and North America have preemptively taken services offline rather than wait. Below is an image showing the hosts potentially vulnerable to CVE-2026-109618 and/or CVE-2026-109619. Source Censys Security firms using honeypots report the same pattern: attackers are scanning many servers and trying to exploit the PHP flaws, but no successful remote code execution has been confirmed in real-world attacks. Sophos says this suggests that attackers are also struggling to exploit the vulnerabilities. Even its own security team needed more effort than expected, despite using AI, to develop a working exploit. However, the risk is growing. VulnCheck researchers have developed an exploit that bypasses ASLR, a security feature that randomizes where code is stored in memory. By combining both vulnerabilities, they can use a memory leak to bypass this protection and run code on the target system. The exploit works, even though there is no evidence of widespread use yet. The attempts that have been caught so far are mostly scanning, fingerprinting, and botnet delivery attempts. Cowrai’s analysis of PHP2Shell traffic in the wild found connections to a Mirai variant called Potassium, which contacts a C2 server, checks for an HTTP 200 response, then downloads a Linux executable as a follow-on payload. “The payload, a likely Mirai botnet variant known as Potassium, was not delivered since the exploitation attempt was unsuccessful due to being blocked by Sophos countermeasures.” states the report published by Sophos. “However, analysis of the sample revealed that is an updated and significantly expanded version of the Mirai-derived botnet variant identified in early 2026.” The malware can download additional files and disrupt networks. It contacts a hard-coded command-and-control server, checks that the server responds successfully, and downloads a file designed to run on Linux. If normal address lookup fails, it can use public DNS servers such as Google DNS and Cloudflare DNS. Like other Mirai malware, it can also launch distributed denial-of-service (DDoS) attacks. Arctic Wolf found that both malware samples had Ethereum Name Service (ENS) resolution disabled, meaning the feature was available but not active. “ENS resolution was disabled in both samples, indicating that the capability was present but inactive.” states Arctic Wolf. “If enabled in future variants, it could complicate disruption by reducing reliance on registrars and DNS providers, requiring defenders to block the ENS names, RPC infrastructure, and resolved command-and-control addresses.” The botnet’s developers have already added the ability to use ENS to find command-and-control (C2) servers. This could help them keep their infrastructure online and make it harder for authorities to take down their domains, but the feature has not yet been enabled. Sophos blocked the attempted Potassium delivery before it succeeded. But analysis of the sample itself showed it’s a significantly expanded version of the Mirai variant spotted earlier in 2026, with added DDoS capabilities and fallback DNS resolution using public servers like 8.8.8.8 and 1.1.1.1 hardcoded as backup. PHP2Shell is being compared to Log4Shell because it affects a huge number of systems. The flaw is in PHP itself, so it can affect websites, content management systems and shared hosting services. Even a basic WordPress site with no extra plugins is at risk if it uses a vulnerable PHP version. Fixes are available in PHP versions 8.2.35, 8.3.36, 8.4.27 and 8.5.12. Older versions, including 7.4, 8.0 and 8.1, need fixes provided by their software distributors or an upgrade to a supported version. Cisco Talos notes that scanning for flaws and using public exploit code is easier than building attacks into custom tools. Advanced attackers may still be developing those tools, so administrators should use this time to patch vulnerable systems. “The true impact of this vulnerability will take time to evaluate.” concludes Talos. “These types of vulnerabilities that impact core technologies tend to cascade into other platforms and applications as they evaluate impact and provide associated patching, harkening back to the waves of exploitation that we saw after Log4J.” Follow me on Twitter: @securityaffairs and Facebook and Mastodon Pierluigi Paganini (SecurityAffairs – hacking, PHP2Shell)
AI agents paying for API calls with the x402 SDKs could sign a payment for their full balance of a supported token if a malicious API asked for it, security firm V12 said on October 22. The cause was the SDKs' default client, which set no spending limit until releases on August 18 added one. The risk was limited to agents who automatically signed payments and whose apps set
Cloud hosting provider Railway published an urgent warning to users after detecting suspicious activity on its systems that store customer environment variables earlier this week. The company says its investigation indicates that, in some cases, an unauthorized third party may have obtained decrypted variables and urges all users to proactively rotate secrets and credentials. “Our investigation to date indicates that an unauthorized third party may have accessed encrypted environment variables and, for some of them, obtained the decrypted contents,” Railway says. “Environment variables may contain secrets such as API keys or database and registry credentials,” the company clarifies. Railway is a cloud platform for developers to deploy and run applications, APIs, databases, and various backend services. It has around 4.8 million users and hosts more than 16 million services.
Japanese gaming giant Capcom has confirmed that hackers breached its membership service, CAPCOM ID, exposing more than 11.7 million customer email addresses and associated information. Apart from emails, exposed data includes account creation date, last login timestamp, number of logins, and the IP address used at the last login. Capcom found that for a small subset of 43 accounts, dates of birth, gender, nicknames, and countries were also exposed. For another subset of 196 accounts, the hackers changed the CAPCOM ID login passwords in an attempt to lock the owners out, but the company took action and reset them. In an alert today, the gaming company says that passwords for all other registered members have not been exposed and that credit card information remained unaffected. Capcom is a Japanese video game developer and publisher known for franchises such as Resident Evil, Monster Hunter, Street Fighter, and Mega Man. The company has 4,000 employees and an annual revenue of $1.24 billion.
Right now, there is probably an AI agent running on a laptop somewhere in your network. It stores API keys and database passwords, can access your internal systems, and works without anyone watching. Nobody reviewed its access. Nobody owns it. If an employee had that much reach with that little oversight, you would call it an insider risk. So why do we still call the agent a productivity tool?
Amazon's Threat Intelligence team on Wednesday said it detected an attempt to exploit the newly disclosed PHP2Shell vulnerabilities dating back to October 17, 2026, against its MadPot honeypot service, as the flaws come under broad exploitation in the wild following the release of the patches. "The exploit attempt originated from an IP address associated with a commercial anonymization network,
Confirm this action.
Leaving now will discard your changes.