The model isn't cooperatingOct 6, 2026
Have you ever felt like a model isn't actually trying to do what you asked? Did you just ask it wrong, or is there something else at play here? In this post, I'll briefly explore this phenomenon, the
UK fashion retailer ASOS confirmed a data breach Tuesday after hackers sent unauthorized push notifications through its mobile app while claiming to have stolen customer data from the company's Snowflake environment. [...]
A new campaign targeting ad account managers uses fake ChatGPT, Gemini, Claude, and Perplexity sites that steal login credentials and multi-factor authentication (MFA) codes through browser-in-browser attacks. [...]
RMM platforms give MSPs privileged access across customer environments, making their security controls critical to limiting risk. Acronis outlines eight controls MSPs should test when evaluating RMM software, from patching and privileged access to recovery and tenant isolation. [...]
It seems that a trend started… I continue my journey discovering more RMM ("Remote Management & Monitoring") tools abused by threat actors! A few days ago, I wrote a diary[1] about ScreenConnect used in the wild. Today, I found another one.
A malicious spreadsheet can make LibreOffice and Apache OpenOffice run an attacker's code as soon as the file is opened, security researchers have shown. There is no warning first, of the kind either program shows before it runs a macro. The attack works only when the program's Java support is enabled. So far, it has only been shown as a proof of concept, and there are no reports of its use in
(c) SANS Internet Storm Center. https://isc.sans.edu Creative Commons Attribution-Noncommercial 3.0 United States License.
The Wikimedia Foundation says rogue OpenAI agents made unauthorized Wikipedia edits and may have been partially responsible for a May outage. [...]
The Wikimedia Foundation, which hosts Wikipedia, has confirmed that it has discovered activity by rogue OpenAI agents on its platforms, including unsuccessful efforts to compromise Etherpad, a public note-taking tool, and edit Wikipedia pages. "The unauthorized bot activities included edits to our wikis, some unsuccessful attempts to exploit a public note-taking tool we host, and heavy traffic,
In 2024, MCP (Model Context Protocol) set out to become the USB-C of AI: one standard for connecting models, agents, and IDEs to tools and data. The protocol delivered. Thousands of developers built servers, and enterprises plugged them into agent workflows. The ecosystem around it fell short. Earlier this year, our team at OX Security, traced critical vulnerabilities in Anthropic's MCP
Over the weekend, Japanese publishing giant Nikkei disclosed that unknown attackers recently breached two employee email accounts and used one to send thousands of phishing emails. [...]
Google has stopped accepting product vulnerability reports through its bug bounty program for its open-source software. The change, in effect since October 1, means researchers can no longer submit security flaws in the code of projects such as Go, Angular, and Protocol Buffers there for a reward. Reports about supply chain compromises are still accepted, and reports filed before October 1 are
A former core infrastructure engineer at an industrial company headquartered in New Jersey was sentenced to 32 months in prison for locking thousands of devices on his employer's network in a ransomware-style attack. [...]
A critical flaw in 8 Atlassian Data Center products, which customers host themselves, allows an attacker with no login access to read specific files in each product's web application root directory. The attacker must already know a file's exact name and path and cannot list what the directory holds. Atlassian disclosed the flaw, CVE-2026-21589, on October 5, rated it 9.3 out of 10, and
The U.S. Federal Bureau of Investigation (FBI) has removed an Accenture contractor for their alleged role in a ShinyHunters-breach that led to the theft of personal details of thousands of bureau employees. That's according to a report from Reuters, citing two sources familiar with the matter. "To date, our review has determined that the incident occurred as the result of a security failure
Unauthorized parties have gained access to the names, addresses, and personal identification numbers of about 8.8 million people, living and dead, in Denmark's national population register, the country's digitalization ministry said on October 5. They used a private Danish company's lawful right to look up records in the Central Person Register (CPR). The ministry has told people never to
A new type of ClickFix attack is using compromised websites to trick users into executing a malicious payload cached in a web browser's cache. "Instead of downloading and executing remote payloads like the typical attack pattern, in this attack, the websites pre-fetch a script payload into the browser cache disguised as a PNG file," the Microsoft Threat Intelligence team said in a post on X.
OpenAI is preparing to add invisible watermarks to text generated by ChatGPT and Codex in the European Union. [...]
Hackers are actively scanning for a Rejetto HFS weak signing key vulnerability, tracked as CVE-2026-61500, that allows session forgery, account takeover, and remote code execution (RCE). [...]
Italy's Data Protection Authority (GPDP) has fined IQVIA €7 million ($7.8M) over poor data-processing practices that the agency says could have put roughly one million patients at risk of data exposure and de-anonymization. [...]