Salesforce Wants Millisecond Decisions Without Paying for a Full LLM
Salesforce introduced Koa Decisions on November 6, the next version of HyperClassifier, which routes work inside Agentforce. In Salesforce's test it returned a choice in 26 milliseconds, against 1.4 seconds for a general-purpose LLM. It handles images and audio, with plans for Flows, Apex and an API. I would ask how often the cheap answer is the wrong one.
Microsoft's Sales Agent Now Has Its Own Mailbox
Sales Development Agent is in Microsoft Frontier public preview, with setup guidance from Scott Magee. It gets its own Entra identity and Exchange mailbox, can be configured in Teams, and connects to Dynamics 365 Sales and Salesforce over MCP. An agent that emails prospects under its own identity is a bigger call than a Copilot sidecar in the CRM.
Cisco Adds Anthropic's Claude Mythos to Red-Team Enterprise AI Agents
Cisco has launched new red teaming, agent observability and vulnerability management features. Its AI Defense platform will add Anthropic’s Claude Mythos to help security teams simulate attacks against AI agents before deployment, Splunk Agent Observability enhancements will link AI token spending to agent performance and Cisco IQ will correlate live threat intelligence with exposed network infrastructure.
Anthropic Targets Enterprise Cyber Risks With Mythos and Fable 5.2
Mythos 5.2 is designed to identify, reproduce, and help fix software vulnerabilities, with access restricted to its Cyber Verification Program. Fable 5.2 brings related advances in coding, reasoning and technical problem-solving, with safeguards intended to limit misuse. Protecting the systems underpinning customer-facing services is essential to safeguarding personal data, service continuity and customer trust.
Wyndham App Flaw Exposes Salesforce Database Containing Customer Data
A security researcher found a hardcoded password in U.S. hotel chain Wyndham’s Android app that exposed an internal Salesforce database with customer names, email addresses, membership numbers and login tokens. The database held millions of entries, though the number of affected records is unclear. Wyndham says it revoked the credential and found no evidence of unauthorized access.