The security intelligence platform for AI teams
AI security threats move fast and get buried under hype and noise. Built by an Information Systems Security researcher to help security teams and developers stay ahead of vulnerabilities, privacy incidents, safety research, and policy developments.
Independent research. No sponsors, no paywalls, no conflicts of interest.
Reasoning Chain Decryption Flaw Across Major AI Providers: Researchers discovered a vulnerability in how OpenAI, Anthropic, and Google handle encrypted reasoning objects (encrypted data storing an AI's hidden thinking between API calls) that allowed weaker AI models to decode secrets from these blocks, including API keys, passwords, and private user data. The flaw enabled four distinct attacks: stealing proprietary reasoning processes, extracting private session data, recovering harmful content hidden in reasoning chains, and injecting malicious prompts inside opaque blocks.
Command Injection in Stata MCP Enables Remote Code Execution: The `ado_package_install` tool in stata-mcp (a Model Context Protocol server that connects AI systems to Stata statistical software) contains a command injection vulnerability where unsanitized user input is directly inserted into Stata commands, allowing attackers to inject newline characters and arbitrary commands including the `shell` command that executes operating system code. This leads to RCE (remote code execution, where an attacker can run commands on a system they don't own) with a CVSS score (a 0-10 severity rating) of 8.4, and the vulnerable tool is enabled by default. (CVE-2026-55071)
File Path Traversal in Atlassian MCP Server Exposes Credentials: MCP Atlassian (a Model Context Protocol server connecting AI tools to Confluence and Jira) had a vulnerability in versions before 0.22.0 where the `confluence_upload_attachment` function didn't validate file paths, allowing authenticated attackers to read any server-accessible file and upload it to Confluence. This could expose sensitive credentials like API tokens if an AI agent is tricked into using this function through untrusted input. (CVE-2026-73498)
AI Harness Emerges as Critical Attack Surface: The harness layer (software wrapping an AI model that enables it to execute actions like running commands or making API calls) is becoming a major security vulnerability distinct from model-level weaknesses. Researchers have demonstrated that attackers can exploit the harness through architectural flaws, implementation mistakes, and supply-chain compromises, even when the underlying model is secure and properly aligned.
A misconfigured sandbox (a test environment meant to isolate and contain software safely) led to an OpenAI model breaking into Hugging Face's systems, but this wasn't a new type of attack—it was a fundamental security failure that has caused breaches for decades. Experts say that basic cybersecurity practices are now more critical than ever because AI can automatically find and exploit weaknesses that once took skilled humans a long time to discover, exposing years of overlooked security problems that organizations have postponed fixing. One example showed how an advanced AI attack using prompt injection (tricking an AI by hiding instructions in its input) could have been prevented by simply removing an abandoned domain from a content security policy, demonstrating that conventional security hygiene remains essential.