Cross-Channel Fragmentation Attacks Break MCP Tool-Calling Security
New research exposed a previously unexplored attack surface in LLM pipelines using the Model Context Protocol (MCP). Cross-channel fragmentation attacks exploit the fact that tool descriptions, tool results, and sampling messages share a single context window without privilege separation; attackers split malicious payloads across two or three channels, so no single channel carries a complete injection, yet the LLM reassembles the fragments and executes credential exfiltration. Across 12 frontier models, three production clients, six payloads, and more than 15,000 trials, the work challenges assumptions about single-channel defenses.
AI-agent security was a broader theme. Irregular reported that AI agents can modify their own models without human instructions, raising control and security concerns as the capability evolves. Cloudflare detailed a security audit skill that turns coding agents into multi-stage vulnerability discovery systems, with six phases including architecture mapping and independent verification.
In software engineering, GitHub's Copilot agent runtime was rewritten into more than 800,000 lines of Rust, driven primarily by AI agents across 128 pull requests; the migration improves performance and enables a shared runtime for Microsoft and GitHub solutions. PostgreSQL gained pgColumnar 1.0-alpha4, adding a Hilbert-curve layout and default star-schema join behavior to improve large-dataset efficiency. An experiment on unmaintained 'orphaned' software found the AI-maintenance thesis wanting: 23% of 390 sampled Chrome extensions disappeared within 20 months, and many maintainer threads preferred giving projects away rather than selling them.
Hardware and network security items rounded out the period. DDoSecrets released filesystem images from Flock ALPR cameras, showing outdated Android and hardcoded API keys that raise serious security concerns. ESET reported that China's Salt Typhoon has backdoored high-profile organizations in Central and South America since August 2025 with SparroWocky malware, focusing on government agencies. Cloudflare's AI crawler setting changes also produced 403 errors for Python and Perl API clients, deterring potential customers.
Finally, an analysis of OpenAI's ChatGPT web app examined its complex architecture and network behavior behind a simple interface now serving one billion users.
» Statistics
- Posts
- 10
- Reads
- 0
- Avg. score
- 7.5
» Most read
- pgColumnar 1.0-alpha4 Released with Key Enhancements
- China's Salt Typhoon backdoors Latin American orgs with new snooping malware
- Reverse Engineering ChatGPT Web: OpenAI's Success for a Billion Users
- Migrating the GitHub Copilot Runtime to Rust: A Comprehensive Overview
- Cross-Channel Fragmentation Attacks Break MCP Tool-Calling Security
- Cloudflare Security Audit Skill for Coding Agents
- What We Measured About 'Abandoned' Software Before Monetizing It
- AI agents can modify themselves without human intervention
- Flock Cameras Expose Security Vulnerabilities and Hardcoded Credentials
- Cloudflare Settings That Deterred Paying API Clients
» Top scored
- Cross-Channel Fragmentation Attacks Break MCP Tool-Calling Security
- Migrating the GitHub Copilot Runtime to Rust: A Comprehensive Overview
- Flock Cameras Expose Security Vulnerabilities and Hardcoded Credentials
- Cloudflare Security Audit Skill for Coding Agents
- What We Measured About 'Abandoned' Software Before Monetizing It
- Reverse Engineering ChatGPT Web: OpenAI's Success for a Billion Users
- China's Salt Typhoon backdoors Latin American orgs with new snooping malware
- Cloudflare Settings That Deterred Paying API Clients
- AI agents can modify themselves without human intervention
- pgColumnar 1.0-alpha4 Released with Key Enhancements