Need to know
- Salesforce acquires customer service AI firm Fin (formerly Intercom) for $3.6 billion, its largest AI deal of the year.
- CrowdStrike ships Continuous Identity for AI Agents, moving from one-time to real-time authorization for every agent action.
- Anthropic remains at odds with the White House over export controls on Fable 5, with talks ending Monday without resolution.
- SailPoint is acquiring AI agent security startup Entro Security for approximately $200 million to bolster its Agentic Fabric platform.
- Cohere reports a surge of inbound enterprise interest as customers scramble for alternatives after the Fable 5 export ban.
New Releases
CrowdStrike launched Continuous Identity for AI Agents inside its Falcon Next-Gen Identity Security platform, replacing one-time authorization with real-time, risk-aware enforcement for every agent action.
- Real-time evaluation checks agent ownership, caller identity, and device risk posture on each individual action, not just at session start.
- Coverage spans on-premises, SaaS, browser, and cloud environments, positioning Falcon as the identity control plane for autonomous AI workloads at scale.
Dataiku announced general availability of Cobuild, an AI building agent that converts plain-language business objectives into governed, production-ready AI projects without code.
- Cobuild enforces enterprise controls from the start of a project rather than as a governance retrofit, directly targeting the compliance gap that stalls AI deployments.
- The launch targets organizations that have modern data foundations but cannot bridge the gap between AI experimentation and auditable production deployment.
Saviynt released Intent-Aware Runtime Authorization (IARA) in its Agent Access Gateway, assessing each AI agent action against identity, context, policy, and declared intent before allowing it to proceed.
- IARA can block and audit actions that fall outside approved limits in real time, creating a verifiable compliance record for every agent decision.
- The release targets the governance gap where AI agents, unlike conventional software accounts, adapt and make independent decisions that can drift beyond their original authorization scope.
Google Cloud introduced the Open Knowledge Format (OKF), a vendor-neutral markdown specification designed to give AI agents curated, structured enterprise context such as table schemas, metric definitions, and join paths.
- OKF is open and model-agnostic, formalizing the pattern of LLM-readable wiki files so enterprises can feed consistent internal context to any agent without rebuilding knowledge stores per vendor.
- The spec ships alongside new data agents across Google's Agentic Data Cloud, targeting the core failure mode where agents stall because they lack access to internal business context.
Funding
Salesforce agreed to acquire Fin (formerly Intercom) for approximately $3.6 billion, betting that a purpose-built customer service AI agent resolving 76% of support volume is a faster path to Agentforce revenue than building the capability internally.
SailPoint is acquiring Israeli AI agent security startup Entro Security for approximately $200 million to add automated agent discovery, inventory, and interaction graphing to its Agentic Fabric platform, signaling that identity governance vendors are racing to own the AI agent visibility layer before it becomes table stakes.
Photon AI closed a $124 million Series A to build vertically integrated AI infrastructure across Africa spanning compute, foundation models, and autonomous cybersecurity, arguing that existing AI architectures structurally disadvantage African users and economies.
Case Studies
Salesforce's acquisition announcement disclosed that Fin's proprietary Apex model resolves an average of 76% of customer support volume end-to-end across chat, email, WhatsApp, SMS, phone, and Slack, outperforming top frontier models on support-specific benchmarks.
- Resolution rate of 76% across live deployments is the concrete production number Salesforce is paying $3.6 billion for, distinguishing Fin from generic LLM wrappers positioned at customer service.
- Apex is purpose-built for customer support rather than adapted from a general frontier model, which Salesforce is presenting as the technical rationale for acquisition over API integration.
Cohere told Bloomberg it is experiencing its fastest inbound enterprise demand after US export controls blocked foreign access to Anthropic's Fable 5 and Mythos 5 models, with customers citing on-premise deployment and sovereign cloud capabilities as primary differentiators.
- On-premise deployment is the structural advantage Cohere is converting: enterprises spooked by overnight model unavailability are treating vendor lock-in to cloud-only API providers as an unacceptable operational risk.
- The Aleph Alpha merger gives Cohere a credible sovereign cloud story in Europe, which is the geography most affected by US export controls on a European-preferred AI provider.
Trending on X
- Fable 5 export ban implications Researchers and policy watchers are debating whether the Trump administration's controls on Anthropic's Fable 5 are a genuine safety measure or competitive favoritism toward OpenAI, with Ethan Mollick and others noting the perverse incentives: if labs can be blocked from using their own models publicly, the rational response is to capture AGI value internally rather than sharing access.
- AI agent harness architecture debate Developers on X are converging on a mental model where an AI agent is essentially a GitHub repo — the LLM is an executable, context is text files, and the harness manages their interaction — framing agent engineering as a software architecture problem rather than a model selection problem.
- Government as AI regulator analogy breaks down A high-engagement thread argues that calls for an FDA-equivalent for AI ignore the combinatorial complexity of model capabilities, and the Fable 5 export control episode is being cited as evidence that blunt government intervention creates market distortion faster than it creates safety.
- Cursor Composer 2.5 Fast benchmark surprise Developers are circulating benchmark data showing Cursor's Composer 2.5 Fast outscores the standard Composer 2.5 on 11 engineering skills while running 32% faster at the same price, prompting debate about whether 'fast' model variants are systematically underrated by buyers who assume speed trades off against quality.
- Anthropic vs. White House standoff escalates Tech Twitter is tracking The Verge and WIRED's simultaneous reporting that export control talks ended Monday without resolution, with community sentiment split between those who see regulatory overreach and those who argue Anthropic's capability rationing approach invited government scrutiny.