Need to know
- Zscaler launched a full zero-trust platform for AI agents, extending Zero Trust Exchange to cover agentic access and activity monitoring.
- Cohere released North Mini Code, a 30B open-weight MoE model targeting enterprise agentic coding workflows.
- Google open-sourced DiffusionGemma, a text diffusion model running 4x faster than standard LLMs on consumer GPUs.
- Salesforce Agentforce hit $1.2B ARR while simultaneously cutting roles in MuleSoft, Marketing Cloud, and parts of the Agentforce unit.
- Claude Fable 5 was jailbroken within hours of launch, leaking its 120,000-character system prompt via multi-agent decomposition.
New Releases
Zscaler today launched a zero-trust platform specifically for agentic AI, extending its Zero Trust Exchange to govern how AI agents connect, access data, and operate on endpoints.
- ZAgent Framework uses natural-language prompts inside Zscaler's Experience Centre to automate SASE configuration, targeting the unmanaged-device and multi-cloud gap.
- Project AI-Guardian expanded to include AWS, CoreWeave, Databricks, Google Cloud, OpenAI, and Saviynt as technology alliance partners alongside existing GSI members.
Cohere released North Mini Code, a 30B-parameter open-weight mixture-of-experts model with only 3B active parameters, aimed squarely at enterprise agentic coding use cases.
- MoE architecture keeps inference cost low by activating 3B parameters per forward pass while maintaining 30B total capacity, making it deployable on-premises without Blackwell-class hardware.
- Open-weight release is a direct competitive signal against GitHub Copilot and Cursor at the model layer, giving enterprises a self-hostable alternative with Cohere's enterprise support wrapper.
Google DeepMind released DiffusionGemma as open source on June 10, a text diffusion model built on Gemma 4 architecture that generates text four times faster than standard LLMs while using less RAM.
- Text diffusion approach refines a noisy token sequence rather than generating left-to-right, enabling parallel decoding and running on consumer RTX GPUs that typically cannot support frontier-class LLMs.
- Nvidia optimization across RTX, RTX PRO, and DGX Spark platforms was announced simultaneously, with Demis Hassabis citing it as a genuine architectural innovation on X.
GitHub removed the waitlist for its Copilot desktop application today, making the agent-native technical preview available to all Pro, Pro+, Max, Business, and Enterprise subscribers.
- Agent-native design spans issue creation through merge, positioning the desktop app as a standalone development environment rather than a VS Code extension wrapper.
- Usage-based billing introduced June 1 now underlies the app, with premium model interactions drawing from monthly credit pools at $0.04 per additional request beyond tier limits.
Funding
Apollo Global and Blackstone finalized a $35 billion private credit package for Anthropic on June 6, structured to fund Google custom TPUs leased to Anthropic, marking one of the largest private credit transactions on record and signaling that AI infrastructure financing has become a mainstream institutional asset class.
Mistral confirmed the acquisition of simulation startup Emmi AI, absorbing more than 30 researchers into its Science and Applied AI teams, a talent-concentration move that narrows the gap between Mistral and larger frontier labs on applied research depth.
Case Studies
Salesforce disclosed that Agentforce has reached $1.2 billion in annual recurring revenue while simultaneously issuing WARN notices for layoffs affecting MuleSoft, Marketing Cloud, and parts of the Agentforce unit itself.
- Revenue concentration in Agentforce is accelerating faster than Salesforce's overall headcount can absorb, with the product hitting $1.2B ARR roughly 18 months after launch.
- Layoffs in the AI unit itself are the telling signal: Salesforce is cutting the human implementation and support layer as agent automation reduces the services wrapper required to deploy the product.
Trending on X
- Claude Fable 5 jailbroken on launch day Researcher Pliny the Liberator defeated Fable 5's safety classifiers within hours using multi-agent decomposition and Unicode tricks, leaking its 120,000-character system prompt, and the AI community is debating whether Anthropic's tiered safety architecture — where flagged queries fall back to Opus rather than refuse — is the right design tradeoff.
- GPT-5.5 vs Fable 5 benchmark rivalry The Berkeley ALE benchmark showing GPT-5.5 beating Claude Fable 5 on autonomous agent tasks touched off a sharp debate about whether safety constraints structurally cap Anthropic models on agentic benchmarks, with Ethan Mollick noting GPT-5.5 Pro matches Fable technically but produces less creative output.
- DiffusionGemma architecture excitement Demis Hassabis's endorsement of DiffusionGemma as a genuine architectural innovation, combined with the 4x speed claim on consumer GPUs, has developers excited about the prospect of running frontier-quality inference locally without data center hardware.
- OpenAI GPT-5.6 imminent release signals Multiple credible sources reporting a GPT-5.6 release this month as a meaningful improvement over 5.5 are fueling speculation that the model cadence has permanently accelerated, with community discussion focusing on whether enterprises can realistically update production deployments at this pace.
- xAI engineer lawsuit over Grok safety Former xAI engineer Devin Kim's lawsuit alleging dismissal after raising Grok safety concerns — timed days before a major SpaceX IPO — is drawing scrutiny from the AI safety community and investor analysts who see the case as a test of whether AI labs face meaningful whistleblower accountability.