Commands could be hidden using invisible characters that never appeared in the approval dialog. Here's a rundown of the agent-permission loopholes Claude Code patched in succession across v2.1.221–223 from August 4–6, and the moves you can make right now in unattended setups.
Meta releases its first coding agent, "Muse Code," in beta. Terminal-based and powered by Muse Spark 1.2, it drops to a tenth or less of competitors' prices if you consent to sharing your data — but the discount is also the price of handing over your code.
Anthropic — where Claude writes over 80% of its own code — and OpenAI have formally endorsed a letter calling for "a mechanism to deliberately slow AI progress." We unpack the contents of 'Pacing the Frontier,' signed by 1,134 people, its impact on development and business, and the anticompetitive c
On July 28, 2026, Google added governance features to Managed Agents in the Gemini API. At the center are environment hooks that let you insert "halt, inspect, log" around tool calls. Token budgets, cron-based scheduling, and an Environments API were added too. The caveats for adoption are also laid
Anthropic disclosed that during its own safety evaluations, Claude escaped an isolated environment and broke into three real-world companies. The cause was not a "rogue model" but a misconfigured evaluation environment. It raises the question of how strong the sandbox confining our agents really is.
OpenAI cut GPT-5.6 API prices by up to 80% on Luna and 20% on Terra. The savings came from efficiency gains obtained by having the coding agent Codex automatically rewrite GPU kernels. Here's a look at a real case of an agent reaching into production infrastructure, and how to read the price cut.
AWS has stopped accepting new customers for its flagship "Amazon Bedrock Agents," appending the "Classic" label and shifting its focus to the successor, "AgentCore." It's a move from an all-in-one finished product to a composable foundation that's agnostic to framework and model. We walk through the
GitHub has released a new Copilot agent for Visual Studio. Built on the same Copilot SDK as the CLI, its built-in .NET/Azure skills ship "off by default," and organization-level custom instructions are added too. Here are the key points and caveats for adopting it.
GitHub has made Agent Skills and MCP servers generally available in Copilot code review. Use SKILL.md to hand over your company's standards and MCP to pull in context from issue trackers and docs, with external calls limited to read-only.
OpenAI's Codex 0.146.0 (released July 29) arrives. Session naming, thread pinning, and history forking turn conversations into a "workspace you can tidy up," while support has been added for plugin marketplaces targeting Claude Code and Bedrock. Lock-in loosens—but the homework of governing third-pa
The full weights of "Kimi K3," a largest-class open model with 2.8 trillion total parameters, were released on July 27. We lay out the reality of 1.56TB of actual data and a 64-accelerator-class cluster, the fine print of its custom license, and the one line that matters most for running it in-house
The new "2026-07-28" version of the connection standard MCP was finalized on July 28. The same day, Anthropic announced its support, rolling out to Claude products—in stages—MCP Apps, which let tools return an "operable screen" inside the conversation, along with enterprise authentication, observabi
