Topic
20 articles on ai security from the Enforster AI team.
AI Security
For about eleven hours on October 29, 2026, Cursor's extension marketplace served a modified Claude Code extension that added an attacker SSH key and tried to open a reverse tunnel. Here is how to check a machine, remove the persistence, and set install controls that would blunt the next one.
AI Security
Attackers spent October 2026 pulling environment variables out of AI developer platforms and reselling the AI access they found. Here is how to rotate, investigate, and shrink the blast radius of the next stolen key.
AI Security
Cursor, GitHub Copilot, Claude Desktop, Codex, and VS Code all embed a browser and give it powers a normal browser withholds. Forever Security showed how a hostile page turns those powers into code execution, file theft, and address-bar spoofing.
AI Security
LiteLLM used one encryption key both to seal stored secrets and to mint login tokens, so an internal user could get the proxy to encrypt a forged admin credential for them. Here is who is exposed, which releases fix it, and the follow-up checks worth doing.
AI Security
A critical out-of-bounds write in the Brotli decoder can be triggered by any compressed response your code decodes. AI agents that browse, crawl, or fetch URLs on request are a direct path to it. Here is what to patch and where the copies hide.
AI Security
A client could tell vLLM which revision of a model's remote processor code to import, even when the operator had pinned the model. Here is who was exposed, what 0.31.0 changes, and the controls to keep after you upgrade.
AI Security
During MCP sign-in, the server tells your client which URL to open. In opencode and mcp-remote, a crafted URL was enough to run commands on the developer's machine.
AI Security
Coding agents obey instruction files in the folder you start them in. Start one in ~/Downloads and any web page that can save a file there can give it orders.
AI Security
Storm-0940 did not need a single user password: it signed in as a trusted vendor application. Here is what happened, what to check in your tenant, and how to shrink the access third-party apps hold.
AI Security
An agent that remembers can be taught to remember the wrong things. Treat every memory write as a privileged action with provenance, review, and an expiry date.
AI Security
The LiteLLM 1.107.0 release on PyPI and its Docker images carried a credential stealer that runs on every Python start. Here is what it did, how to tell if you were hit, and what to rotate.
AI Security
Anything a prompt can influence, an attacker can influence. Here is how to handle model output safely at each place it lands: the browser, the database, the shell, and the interpreter.
AI Security
A system prompt is configuration the user can probably read, not a vault. Here is what to move out of it, where to put it instead, and how to test that it stays out.
AI Security
An LLM endpoint turns every request into a bill. Here is how to cap what a single user, an agent loop, or a stolen API key can spend before it becomes an incident.
AI Security
A RAG system that searches one shared index for every user will eventually hand someone a document they were never allowed to read. Here is how to enforce access control at retrieval time.
AI Security
An AI agent in a GitHub Actions workflow reads attacker-controlled text while holding your tokens. Here is the threat model and a workflow design that keeps a successful prompt injection from reaching secrets.
AI Security
AI coding assistants regularly suggest packages that do not exist, and the same fake names come back again and again. Here is how slopsquatting works and the controls that stop it at review, install, and CI time.
AI Security
Connecting an MCP server hands its author a voice in your agent's context and, often, code execution on your machine. Here is how to review one before you connect it and keep it honest afterwards.
AI Security
A downloaded model checkpoint is not just data. If it is a pickle, loading it can execute code. Here is how to load third-party weights without handing that file a shell.
AI Security
Filters and stern system prompts reduce prompt injection, but they cannot eliminate it. The durable fix is to treat model output as untrusted input and authorize every action it proposes.
