The bug was simple: the login let you choose where your sign-in code went. Proving it on a live website was the hard part. TL;DR During an authorized assessment, an AI pentesting agent found a critical authentication bypass on a major insurer's production website, on its own. One login request let the user choose where the one-time passcode went. The attack was the easy part. Proving it
A critical heap buffer overflow vulnerability (CVE-2026-39919) in the Ghostscript interpreter and rendering engine could allow remote code execution through specially crafted PDF files. The flaw was discovered on August 23 by researchers at Defiant's platform Wordfence using their Argus AI-powered system. Ghostscript is a PostScript and PDF interpreter that can render PDFs as images, generate thumbnails for previews, and optimize PDF files by compressing embedded images and removing unnecessary font data. When generating a PDF thumbnail, WordPress can use the PHP Imagick extension to call ImageMagick, which launches Ghostscript. A specially crafted PDF file containing a JPEG 2000 image can trigger incorrect handling of low-bit-depth samples and mismatched component resolutions. Using Argus, the researchers developed two exploits for CVE-2026-39919 that force Ghostscript to write 1,024 bytes into a 256-byte buffer, corrupting nearby memory and leading to unauthenticated remote code execution in a WordPress deployment.
The U.K. Housing Ombudsman Service is investigating a third-party IT security incident that impacted its case management system on Tuesday. Established under the Housing Act 1996, the Housing Ombudsman is an executive non-departmental public body sponsored by the Ministry of Housing, Communities, and Local Government. The watchdog settles disputes between social housing residents and social landlords in England, including housing associations, landlords, managers, and letting agents. On Tuesday, the service added an alert to its official website, warning that it's working with the affected service provider to assess the incident's impact and said social housing residents cannot file new complaints until systems are back online. Those who must raise new complaints are advised to call the Housing Ombudsman's phone line or email the watchdog at info@housing-ombudsman.org.uk.
A flaw in SAP's Cloud SDK for Python can send an application's OAuth2 client credentials to a server an attacker controls, SAP said in a security advisory on October 13. The flaw affects multi-tenant applications on SAP Business Technology Platform (BTP) that use the SDK, SAP's official Python client for BTP services. An attacker must be able to control the tenant subdomain,
Attackers continue exploiting gaps in MFA coverage, using stolen credentials to access VPNs, RDP, and Windows systems that still rely on passwords alone. Specops explains why protecting cloud applications isn't enough and how organizations can close these authentication gaps.
SAP has addressed 20 security flaws across multiple products in its October 2026 security updates, including a maximum-severity vulnerability in the Cloud SDK for Python. The SAP Python SDK is the official programmatic toolkit for developing AI agents and Python applications on the company's SAP Business Technology Platform (BTP) services. Tracked as CVE-2026-89111, the maximum-severity vulnerability patched this month can be exploited by remote attackers without privileges to steal credentials. Customers who can't immediately deploy the October 2026 SAP patches can temporarily mitigate the issue by using an allowlist to validate tenant_subdomain values. "The SAP Cloud SDK for Python contains an OAuth2 SSRF vulnerability affecting multi-tenant BTP applications. An attacker who controls the tenant_subdomain parameter can redirect OAuth2 token requests to an attacker-controlled server, causing client credentials to be exfiltrated," according to a Tuesday advisory.
GitLab fixed a critical vulnerability with a maximum CVSS score of 10.0 that could let unauthenticated attackers run code on its servers. GitLab has addressed a critical security flaw, tracked as CVE-2026-97229 (CVSS score of 10.0), that could let attackers run code on its servers without logging in. The flaw affects Community Edition (CE) and Enterprise Edition (EE), it is an integer overflow bug caused by weak validation of regular expressions in CI/CD configurations. The issue affects GitLab versions 19.2 before 19.2.8, 19.3 before 19.3.5, and 19.4 before 19.4.2. Users should upgrade to a patched version to protect their systems. GitLab advisory “GitLab has remediated an issue that, under certain conditions, could have allowed an unauthenticated user to execute arbitrary code on the GitLab server via a specially crafted regular expression in a CI/CD configuration, due to insufficient regular expression validation.” reads the advisory. The researchers joaxcar reported this vulnerability through the company’s HackerOne bug bounty program. The company also addressed these flaws: CVE-2026-92883 (CVSS score: 8.7) – A security flaw where an attacker can trick another user’s web browser into running malicious JavaScript code through a specially designed repository. CVE-2026-16301 (CVSS score: 8.5) – An authorization flaw that lets a user with a Developer role bypass security permissions and run automated testing/deployment (CI/CD) pipelines on protected code branches. CVE-2026-14814 (CVSS score: 8.2) – A pattern-matching flaw that allows an attacker to manipulate cached files, causing target users’ automated pipelines to run malicious code and potentially leak sensitive environment variables. In early October, GitLab released patches for a critical vulnerability in its AI Gateway, tracked as CVE-2026-90970 (CVSS score of 9.9), that could allow an authenticated user with access to the Duo Agent Platform to execute arbitrary commands on the gateway. GitLab disclosed the flaw on October 2 and fixed it with the release of AI Gateway versions 19.2.4, 19.3.2 and 19.4.1. The issue affects the way the AI Gateway handles custom flow prompt templates. According to GitLab, a user with Duo Agent Platform access could use a specially crafted flow configuration to escape the prompt template sandbox and execute commands on the AI Gateway host. “GitLab has remediated an issue in the GitLab AI Gateway that, under certain conditions, could have allowed an authenticated user with Duo Agent Platform access to escape the prompt template sandbox via a specially crafted flow configuration, leading to arbitrary command execution on the AI Gateway.” reads the advisory. The key point is that the attack does not require an unauthenticated connection. GitLab’s advisory does not provide further details about the exact conditions required to exploit the flaw or the permissions needed beyond that access. GitLab credited the HackerOne researcher invisiblemeerkat with responsibly reporting the vulnerability. Follow me on Twitter: @securityaffairs and Facebook and Mastodon Pierluigi Paganini (SecurityAffairs – hacking, CVE-2026-97229)
A user opens a browser tab, pastes data into a chat assistant, and nothing is installed. A developer adds a coding assistant as an editor extension, and nothing shows up in the installed programs list. Someone downloads a local model runtime, a program that runs models directly on the machine instead of sending data to an external service, because they were told not to send company data
Anthropic announced that its internal AI models under “high-risk and cybersecurity evaluations” no longer have internet access following incidents involving U.S. government websites at the federal, state, and local levels. The instances leading to the decision include AI models that exploited website vulnerabilities, submitted forms, bypassed data restrictions, and circumvented safeguards. However, the company said that the misaligned behavior had a minimal real-world impact and assessed the behavior to be less severe than previously reported incidents. The AI company notified the White House about the unauthorized activity: Anthropic estimates that the agents may have been motivated by overly generous rewards during reinforcement learning, which may have encouraged them to find unauthorized workarounds rather than stop when a task was blocked or impossible.
CyberXero used 51 AI agents to support cyberattacks, stealing 628,000 Ukrainian records, including sensitive data from energy infrastructure. CyberXero is an operation that combines traditional hacking tools with AI agents to carry out cyberattacks. According to SOCRadar researchers, the attackers use the same infrastructure for two different campaigns: one targets WordPress websites and e-commerce platforms worldwide, while the other focuses on Ukrainian energy and critical infrastructure. The actor behind CyberXero is a Russian-speaking, financially motivated threat actor. This type of attacker, known as an initial access broker, breaks into networks and sells access to other criminals instead of using it directly. One campaign focuses on attacking as many targets as possible, while the other involves carefully selected targets in Ukraine. “SOCRadar’s Threat Research Unit (STRU) has documented CyberXero, a Russian-speaking, financially motivated Initial Access Broker that pairs commodity offensive tooling with an AI orchestration layer running on its own infrastructure.” reads the report published by SOCRadar. “The operation runs two pipelines at once: a global, automated campaign against WordPress and e-commerce platforms, and a curated, manual campaign against Ukrainian energy and critical infrastructure. More than 628,000 Ukrainian individuals have confirmed data in the actor’s possession, including residents of Kharkiv, a city on an active war front.” The investigation only happened because the threat actor made one classic mistake, an open directory with zero access control sitting on a server. That single misconfiguration exposed the actor’s entire working directory, more than 90,000 files, including AI session logs, scripts with plaintext tokens, and already-stolen victim data. The exposed server was found at “46.21.250[.]135,” hosting a live Linux workstation’s complete contents. Session logs recording every planned target, every refused task, and every workaround attempt are, in practice, the offensive equivalent of a keylogger running against the attacker’s own operation. The global pipeline runs on a custom tool called wp2shell, which abuses a flaw in the WordPress REST API to inject SQL, plant a rogue admin account, and drop a webshell. The scale is genuinely industrial. A single automated run scanned 4,708 targets, confirmed access to 429 WordPress admin panels, and deployed 32 webshells, all within 61 seconds. That’s not a person clicking through targets one at a time, that’s infrastructure built to move at machine speed. The target list for this side of the operation exceeds 56,000 identified URLs, spanning e-commerce platforms and WordPress sites with no particular country or sector pattern. The second campaign is more targeted. The attacker manually selected targets in Ukraine, focusing on energy and utility companies, including the country’s national electricity transmission operator. Researchers confirmed data theft from four Ukrainian organizations. One of the most serious cases involved a district heating provider in Kharkiv, a city close to the front line. A password stored directly in the attacker’s script gave access to the provider’s systems. An automated tool then extracted more than 564,000 customer records, including home addresses and account numbers. SOCRadar warns that this data could put people at physical risk because Kharkiv is in an active war zone. Stolen customer information is usually a privacy concern, but in this situation, linking real names to home addresses could create much more serious dangers. “Because Kharkiv sits on an active conflict front, this dataset poses a potential physical risk to identifiable civilians that goes well beyond a conventional data breach.” continues the report. “The most elaborate chain, against an e-commerce portal, combined blind SQL injection, a time-constrained password-reset token extraction, and PHP deserialization to reach remote code execution and full database access.” The attacks stand out from typical cyber operations because the actor ran up to 51 specialized Claude Code agents on the primary workstation, assigned to roles like exploit chaining, lateral movement, and reconnaissance, plus a separate AI pentesting framework wired directly into a Cobalt Strike command server. SOCRadar’s researchers found the actor had also built a reusable script to get around model safety refusals, framing every request as an authorized penetration test on infrastructure the actor supposedly owned. That trick isn’t a technical bypass, it’s social engineering aimed at the model instead of a person. And it didn’t always work. The logs show the model declining real requests multiple times, including four separate refusals to install a backdoor, disable a firewall, move laterally, or deploy a webshell. Safety measures held in plenty of individual sessions. The actor’s workaround was just opening a new session each time, since a refusal doesn’t carry forward once the context resets. Researchers identified the operator by combining several clues: a paid AI subscription linked to a personal email address, a forum post using the same online alias, and activity that consistently matched Moscow time. The campaign affected more than 40 organizations across Ukraine, Poland, China and Pakistan. The main lesson for defenders is to protect AI agent session logs as carefully as passwords and other sensitive credentials. These logs can reveal an attacker’s plans, the systems they accessed, what worked and where their attempts failed. If exposed, they can provide a detailed picture of the entire operation. The report includes Indicators of Compromise (IoCs) for this campaign along with detection and hunting guidance. Follow me on Twitter: @securityaffairs and Facebook and Mastodon Pierluigi Paganini (SecurityAffairs – hacking, CyberXero)
Confirm this action.
Leaving now will discard your changes.