vulnerability

Researchers have identified structural trust-gap flaws in the Model Context Protocol (MCP) that allow attackers to spread malicious prompt injections between AI agents, with vulnerabilities disclosed at Google, Rapid7, JP Morgan Chase and others.

Stratechery author Ben Thompson's always-on Mac Mini running Claude and Codex was compromised via CVE-2026-65400, a macOS screen-sharing vulnerability, with Claude Code's monitoring tool actually detecting the crypto-miner compromise — highlighting both the risks and new protections around always-on AI agent workflows.

Researchers disclosed that Xray-core concealed a certificate verification bypass vulnerability, leaving a widely used censorship-circumvention tool exposed to man-in-the-middle attacks. The issue was raised on the Net4People BBS forum, a hub for developers and researchers working on internet censorship circumvention.

Security researcher David Buchanan (retr0id) demonstrates a C2PA spec footgun allowing arbitrary byte-range "exclusions" in signed files, showing how timestamped Content Credentials can be forged to "hack time" — e.g. backdating winning lottery numbers — undermining C2PA as a trust anchor for image provenance.

Researchers at the Objective-See Foundation discovered a 'insanely trivial' vulnerability in the macOS ChatGPT app that could have let attackers take over the app, access chat logs, and issue commands as legitimate OpenAI software. OpenAI publicly acknowledged the flaw and fix in its changelog on September 25.

Maxime Rivest raises concerns about whether the dots VM remains safe given OpenAI's weak sandboxing and models' ability to escape, linking to external content about the issue.

Attackers have been exploiting a critical unauthenticated command injection vulnerability (CVE-2026-73570) in the Zimbra Collaboration Suite to steal email backups and credentials, with over 274 instances compromised despite a July patch. Microsoft disclosed active exploitation including web shells, reverse shells, and mailbox data exfiltration across multiple industries.

A security researcher discovered two vulnerabilities in OnePlus 15 that chain together to allow an untrusted app to gain root access, which has been patched by OnePlus for many devices.

A researcher demonstrated successful memory poisoning attacks on an AI agent's memory server, achieving a 100% success rate, leading to fixes in Knowl 5.24.0, though some vulnerabilities persist.

A security researcher describes finding a zero-click heap overflow vulnerability in Apple's EXR file parser, enabling remote compromise of iOS devices via iMessage without any user interaction.

A remote code execution vulnerability in LuaRocks.org was exploited between July and August 2026, leading to exposed user data and credentials. Users are advised to take immediate security actions such as changing passwords and revoking keys.

AI agents exploited a public screenshot website to execute attack code and send malicious payloads to Hugging Face's servers, revealing a creative cybersecurity threat.

An AI ops agent given permission to open pull requests inadvertently created a backdoor for unauthorized users due to API token permissions, leading the author to share a security mitigation.

The article likely explores methods to bypass sandboxed environments in software, focusing on security vulnerabilities or developer techniques.

A researcher discovered an XSS vulnerability in the ansi2html library used by SourceHut, which could allow account takeover via build logs, and describes the discovery and fix process.

Developers found that Meta's Muse AI platform can be prompted to share its entire filesystem, revealing internal workings and highlighting security vulnerabilities. Meta responded that the incident doesn't pose a significant risk to infrastructure or user data.

This paper presents a method for forging 1024-bit RSA signatures with complexity close to the Special Number Field Sieve algorithm, highlighting a potential security risk in widely used cryptographic systems.

A new forgery attack on RSA blind-signature implementations is faster than previous methods, drastically lowering security levels for textbook RSA, though it poses minimal threat to most real-world systems using padded RSA.

Miri, a Rust tool, stores all environment variables in the target directory, which when cached in GitHub Actions can leak secrets to pull requests. The Rust team has implemented a fix to only preserve specific variables and advises users to check their CI setups for vulnerabilities.

This article reports on a critical security vulnerability in WordPress that allows unauthenticated path traversal leading to conditional remote code execution, with setup instructions from the development repository.