Anthropic launched Claude Sonnet 5 on June 30 as the new default for Free and Pro tiers, replacing Sonnet 4.6 with a model that closes the gap to Opus 4.8 at a fraction of the cost.

• Introductory pricing is $2/M input tokens and $10/M output tokens, undercutting Opus 4.8, GPT-5.5, and Gemini 3.1 Pro, which changes the unit economics for teams running multi-step agent workflows at scale.
• Agentic autonomy is the headline design goal: Sonnet 5 can plan, browse, and execute terminal commands at a level that until now required larger flagship models, with built-in cyber safety classifiers replacing the controls that caused Fable 5's 18-day restriction.

BeyondTrust launched a private beta of AI Agent Security, a Pathfinder module that applies existing endpoint privilege rules to AI coding tools including Claude Code, Cursor, Microsoft Copilot, and OpenAI Codex.

• The core extension treats AI coding agents as first-class software actors subject to the same privileged-action controls as human users, meaning an agent cannot escalate beyond the permissions the employee it operates as already holds.
• Scope at launch covers the most widely deployed enterprise AI tools, positioning BeyondTrust to capture governance spend before organizations standardize on a dedicated non-human identity framework.

Jamf announced general availability of AI Governance, integrated into Jamf for Mac, giving IT and security teams visibility, control, and compliance audit trails over AI tools running across managed Apple endpoints.

• The gap it closes: most enterprises can see which SaaS apps are installed but cannot audit which AI tools are active, what data they access, or whether usage meets policy, a problem Jamf frames as structural rather than configuration-specific.
• Native integration means no additional agent or third-party connector is required, lowering deployment friction for organizations already standardized on Jamf for zero-touch provisioning.

Microsoft moved Agent 365 to general availability, giving enterprise admins a centralized control plane with identity, access, and monitoring controls over AI agents as ungoverned shadow AI becomes a security risk.

• What it governs: any agent running within Microsoft 365 tenants, with policies that can restrict which agents employees can deploy and what data those agents can reach, addressing the gap between Copilot Studio proliferation and IT visibility.
• Timing signal: the GA follows a documented rise in employees running unsanctioned agents on corporate data, a vector that existing DLP and CASB tools were not designed to intercept.

In a head-to-head benchmark against a leading general-purpose LLM, Archer's purpose-built Evolv model achieved 95% verified accuracy on regulatory change management with zero errors on critical dates, while the general LLM was confidently wrong 35% of the time.

• Speed and cost: Archer Evolv completed regulatory change tasks 80x faster and at 92% lower cost than the general-purpose LLM it was compared against, metrics that directly affect compliance team staffing and audit cycle time.
• The confidence-error problem: the benchmark specifically exposed a failure mode where general LLMs return wrong regulatory dates with high confidence, a silent error that flows into compliance calendars undetected and creates missed-deadline liability.

Palo Alto Networks and CrowdStrike both reported their best quarters ever, with respective stock rallies of 95% and 113%, driven by enterprise demand for AI-native security as AI agents outnumber human users on corporate networks.

• Demand driver: both companies cited identity security as the fastest-growing segment within their platforms, reflecting the expansion of non-human identities from AI agents that need the same access governance as employees.
• SentinelOne confirmation: SentinelOne stock rose 12.2% on the same day after announcing agentic AI security integrations, suggesting the AI-threat tailwind is lifting the entire endpoint and identity security category simultaneously.