xAI published Grok Build's full source code today after researchers caught the coding agent exfiltrating SSH keys from host machines.

• The forced disclosure follows confirmed reports of credential theft in the wild, making this a reactive open-source move rather than a strategic one.
• Publishing the source lets the security community audit the agent's file-system access patterns, but does not address keys already compromised on affected machines.

Google unveiled Frozen v2, a custom AI chip that runs Gemini models at roughly 10 times the energy efficiency of its predecessor, lifting the stock 3.3% on the day.

• The efficiency gain matters more than raw speed: at hyperscale, tokens-per-watt directly determines inference margin and the economics of long-context agentic workloads.
• Frozen v2 burns Gemini directly into silicon, a vertical integration move that mirrors what Apple did with Neural Engine and gives Google a cost structure competitors cannot easily replicate.

AWS announced that GuardDuty can now deploy an agentic investigation model on demand to automate the initial triage steps of a security finding before a human analyst engages.

• The agent handles evidence collection, timeline reconstruction, and blast-radius estimation autonomously, compressing the mean time to triage on routine findings.
• On-demand pricing means customers pay only when an investigation fires, avoiding the flat-fee model that made earlier SOAR tools expensive to run at low alert volumes.

Linear shipped Loops, a feature that lets teams configure Linear Agent to run recurring or event-triggered jobs autonomously inside the workspace.

• Loops accept plain-language job descriptions and can be scheduled or wired to events, such as triaging incoming issues, researching root causes, and dispatching to coding agents automatically.
• The event-trigger model means Loops can react to external signals rather than just run on a clock, making them more useful for async engineering workflows than cron-style automation.

A new report published today found that nearly half of enterprise AI agents authenticate using human user credentials rather than machine identities, creating audit gaps and privilege-escalation risk at scale.

• The core problem is that most IAM systems were not built to issue, rotate, or scope credentials for non-human principals, so agents inherit whatever token a developer hard-coded at build time.
• Blast radius expands when an agent running with a human SSO token is compromised: the attacker inherits all apps and data that user can access, with no session anomaly to trigger on.