OpenAI shipped Lockdown Mode to all ChatGPT tiers today, disabling web browsing and deep research to cut the attack surface for prompt injection.

• Available to all plans: Free, Plus, Pro, and self-serve Business accounts can enable the toggle, making it the first prompt-injection control at ChatGPT's scale.
• Trade-off is explicit: Lockdown Mode sacrifices live internet access entirely, a design choice that signals OpenAI sees agent tool-use as the primary injection vector.

Cursor's second in-house coding model, Composer 2.0, surpasses Claude Opus 4.6 on Terminal-Bench 2.0 while pricing input tokens at $0.50 per million versus Anthropic's $5.

• Benchmark leadership: Composer 2.0 beats Opus 4.6 and GPT-5.4 on Terminal-Bench 2.0, the first time Cursor's own model has topped third-party frontier models on a coding benchmark.
• Pricing arbitrage: At 10x lower cost than Opus 4.6, Cursor eliminates the core argument for using Claude or Codex as the backend inside its own IDE.

Silverfort today launched an integration that evaluates every access request a Microsoft Copilot Studio agent makes in real time, applying conditional access and MFA before the action completes.

• Runtime enforcement: Unlike post-hoc audit logging, Silverfort's controls intercept the agent's identity at the moment it requests access, treating the agent as a non-human principal subject to the same Zero Trust policies as users.
• Gap it closes: Copilot Studio agents previously inherited broad service account permissions; this integration scopes each action to task-specific, short-lived credentials.

Salesforce announced Agentforce Marketing at its Connections 2026 event, adding agents that autonomously qualify leads, generate content, and run campaigns inside existing CRM data.

• Trust gap is the headline risk: MIT research presented at Connections found 81% of marketing professionals trust agentic AI with customer data, but only 36% of consumers agree — a gap Salesforce has not yet addressed with disclosures or controls.
• Positioning: The launch targets the administrative overhead reduction argument, not net-new pipeline, which is where most Agentforce ROI cases have been weakest.

Anthropic is paying roughly 1,000 external engineers $280 per task to guide Claude Code through high-quality coding workflows, with the program now responsible for the majority of code quality improvements in Claude Code.

• Humans setting the ceiling: Engineers do not write code themselves but instead direct Claude Code step by step, with their judgment on what constitutes a good result becoming the training signal.
• Cost signal: At $280 per task and 1,000 contractors, this is a multi-million-dollar ongoing investment in human-in-the-loop RLHF specifically for agentic coding, not generic preference data.

Tenable announced Claude-powered agentic workflows inside Tenable One's Hexa AI engine, using Anthropic's models to prioritize and automate remediation across the full attack surface.

• Workflow specifics: Claude handles multi-step reasoning across fragmented security telemetry to surface which exposures represent actual risk paths, not just raw CVE counts.
• Ecosystem context: Tenable also launched the One Open Connector today, letting third-party data flow into the same platform, giving Claude broader context than any single vendor feed.