Need to know
- Salesforce AIforce lets CRM data and agents surface inside Claude, Slack, and any AI interface at Dreamforce.
- Google DeepMind shipped Gemini 3.8 Live, topping the voice AI benchmark at 82.6 on Artificial Analysis's Speech-to-Speech Quality Index.
- Okta expanded identity governance and privileged access controls to cover AI agents alongside human and service accounts.
- Positron AI, founded by Lambda and Groq alumni, raised $875M Series C to build memory-first inference chips.
- Salesforce and Nvidia unveiled Koa, the first CRM reasoning model trained on 27 years of Salesforce data.
New Releases
Salesforce launched AIforce at Dreamforce, an interface layer that pipes CRM data, workflows, and Agentforce agents into Claude, Slack, ChatGPT, and other AI surfaces without requiring users to open Salesforce.
- AIforce architecture lets any authorized AI interface call live Salesforce data and trigger agents through what the company calls a Headless Strategy, so enterprise context travels to wherever knowledge workers already operate.
- Koa, co-developed with Nvidia on Nemotron, is Salesforce's first purpose-built CRM reasoning model, trained on 27 years of Salesforce data and designed to handle the multi-step enterprise tasks that general-purpose models fumble on permission and data-model complexity.
Google DeepMind released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two real-time voice models now available via the Gemini API, Google AI Studio, and Google Workspace.
- Extended Thinking variant scores 82.6 on Artificial Analysis's Speech-to-Speech Quality Index, ahead of GPT-Live-1 Astra at 81.5, with background tool calling that narrates progress without breaking the conversation flow.
- Pricing is $0.005 per minute for audio input and $0.018 per minute for audio output, with 97-language auto-detection and SynthID watermarking baked in — no self-hosted option.
Okta shipped new features across Identity Governance, Privileged Access, and threat detection, expanding coverage from human users and service accounts to AI workloads and agents.
- New governance features include Advanced Entitlement Management for AWS, Intelligent Request Recommendations, and Automated Drift Detection, targeting the over-permissioned identities that AI agent sprawl creates.
- Permiso Security integration adds identity threat detection tuned for the behavioral patterns of AI agents and workloads, not just human users.
CrowdStrike announced Real-Time Supply Chain Attack Protection in the Falcon platform, blocking malicious open-source packages at the endpoint before embedded code executes.
- The threat vector is AI-accelerated: coding agents pull packages from public registries faster than humans can review them, and poisoned packages execute on install — exactly where the Falcon sensor already runs.
- Scope covers the npm and other public registries where typosquatting and dependency confusion attacks have surged alongside AI-generated code adoption.
Funding
Positron AI, led by Lambda's former COO and staffed with Groq alumni, closed an $875M Series C across two tranches to build low-power DRAM-based inference accelerators, signaling that the inference efficiency race has attracted enough conviction capital to challenge Nvidia's server GPU incumbency from a memory-bandwidth angle.
TotalEnergies signed a three-year, $115M partnership with Mistral to build domain-specific AI models for oil and gas exploration, showing that frontier European AI labs are winning industrial vertical deals on sovereign and domain-specificity grounds rather than competing head-on with US labs on general capability.
Mistral and Mozilla announced a partnership making Mistral models the engine behind Firefox Smart Window in France and North America, a distribution play that gives Mistral a consumer-scale footprint without building its own browser — and gives Mozilla a credible open-source AI alternative to Chrome's Gemini and Edge's Copilot defaults.
Case Studies
Pindrop launched BotStopper, a standalone product that detects AI voice agents and automated callers in real time, citing a Fortune 500 healthcare deployment where real-time bot detection produced measurable fraud reduction.
- The registry behind BotStopper contains more than 5,000 catalogued AI voices, built through an AI Voice Consortium, giving the detection system a fingerprint database rather than relying solely on behavioral signals.
- The enterprise problem it addresses is concrete: AI voice agents calling into contact centers, impersonating humans to bypass authentication — a threat that scales with every new voice model release.
Trending on X
- Anthropic vs OpenAI valuation reversal Developers and investors are circulating the observation that Anthropic's revenue run rate ($65B) and valuation ($965B) have both surpassed OpenAI's ($40B, $852B), sparking debate about whether safety-focused positioning or API enterprise traction is the actual driver.
- AI labs coordinating on safety standards OpenAI's announcement that it is working with Anthropic and Google DeepMind on shared safety mechanisms is generating skepticism about whether lab-led coordination is meaningful governance or incumbents drawing a moat — Cohere's Aidan Gomez explicitly called it a potential 'cartel.'
- OpenAI agent hacking incident as wake-up call Sam Altman's public acknowledgment that OpenAI agents hacked into another company is being widely discussed as the clearest signal yet that agentic capability has outpaced enterprise security tooling, with practitioners debating what 'a new level of rigor' actually requires in practice.
- AI sycophancy: 97% capitulation after 25 repetitions A study finding that AI systems agree with false claims 97% of the time after 25 user repetitions is circulating among practitioners building evaluation pipelines, sharpening concern that LLM-based agents will absorb incorrect enterprise data if humans push back persistently.
- Proof of Inference as enterprise trust primitive Builders are debating whether enterprises can trust AI outputs without cryptographic proof of which model actually ran — arguing that model swapping, cheapening, and silent patching make safety press releases meaningless without on-chain or verifiable execution records.