Stopping rogue AI agents requires more than a kill switch
Disabling an AI model does not always stop its workflows. Teams must revoke credentials and restrict network paths to contain unexpected agent behavior.
AI & LLMs
Models, tooling and practical AI features in business software.
Anthropic expands Claude connectors to open full enterprise apps inside chat
Anthropic announced expanded interactive connectors allowing Claude to open and edit full applications like Adobe and HubSpot directly within the chat interface, respecting existing user permissions.
Anthropic launches Mythos 5.2 and Fable 5.2 with lower costs and RSI concerns
Anthropic released Mythos 5.2 and Fable 5.2, cutting Fable input costs to $8 per million tokens while addressing recursive self-improvement risks.
Hidden text in PDFs can trick AI document processors
A QA engineer demonstrates how hidden text and layout tricks in PDFs can inject false data into AI systems, proposing a multi-layered defense strategy.
Local LLM benchmark reveals extraction flaws in knowledge graph building
A new reproducible leaderboard tests 22 local models on knowledge graph extraction and chat, showing that model size does not guarantee accuracy.
Self-hosted AI gateway outage reveals credential handling flaws
A developer shares how a self-hosted model gateway triggered a vendor block due to shared OAuth tokens, leading to a redesign of task dispatch and error handling.
Llama-server Responses API falsely reports truncated outputs as completed
A bug in llama.cpp build b11753 causes the /v1/responses endpoint to mark cut-off replies and invalid tool calls as completed, risking data corruption in self-hosted AI agents.
Google releases Gemini 4 Argon with unified token budget and new pricing
Google has made Gemini 4 Argon available to developers, introducing a shared 1 million-token limit for input and output alongside introductory pricing that matches key competitors.
OpenCode 2.0 unifies terminal, web, and desktop interfaces under one service
Anomaly releases OpenCode 2.0, featuring a shared background service, rebuilt API, and Code Mode for efficient tool integration.
Personal AI agents are accessing work systems via employee identities
Employees are connecting personal AI assistants to corporate tools through OAuth and saved credentials, creating new identity security challenges for IT teams.
Distinguish truncation from errors in nightly LLM evaluations
A new evaluation pattern separates transport failures and token limits from actual model mistakes to prevent misleading quality metrics.
AI agents inherit full user permissions, breaking audit trails and security boundaries
A security analysis reveals that AI coding agents borrowing human credentials gain excessive access and leave no distinct audit trail, creating significant risks for self-hosted infrastructure.
TinyWebUI offers a local-first chat interface with built-in MCP tools
TinySuiteHQ released TinyWebUI, a self-hosted chat workspace that runs locally with SQLite storage and includes pre-configured search and memory tools.
Docsy moves to Linux Foundation to optimize docs for AI agents
Google’s Docsy documentation theme joins the Linux Foundation, adding features like llms.txt to help AI agents consume technical content more effectively.
Self-hosted AI agent platform for GitLab reaches 133 releases
The langgraph-harness project, a self-hosted AI coding agent for GitLab, has reached 133 releases with a 67% comment acceptance rate.
