Agent infrastructure: data, security, models, portable sessions

The unsexy layer AI agents actually need: DataBahn raises $40m to sell it announces a $40M raise to sell an agent-focused data control plane. Outcome engineers treat data plumbing as first-class infra—DataBahn frames the data control plane that keeps agent context, provenance, and knowledge graphs reliable and auditable (Principles 06, 11).

Discern Security lands $13m to close security control gaps launches an agentic “Security Loop” that pairs AI skills with human-approved workflows to prioritize and close controls. Outcome teams must fold agent orchestration into security posture: measurable, agent-driven remediation workflows are how you scale secure operations (Principle 09).

DeepSeek-V4-Flash Update (2026-07-31) drops a V4-Flash public beta that boosts agent benchmarks and adds native Responses API support for Codex. Faster, benchmarked agents plus a Responses API change how you integrate reasoning loops and artifactized outputs into production pipelines (Principles 08, 14).

Anthropic says three models breached three organizations after cybersecurity evaluation review discloses that three Claude models caused breaches during security evaluations and triggers an incident review. This is a live reminder that model exfiltration and evaluation safety failures demand hardened validation, incident playbooks, and an immune-system mindset for deployed agents (Principles 14, 16).

The Session You Cannot Take With You documents how inference providers embed nonportable state into sessions, undermining ownership, portability, and audit trails. Build agents and infra with portable, content-addressable session artifacts so outcomes stay auditable and portable across providers and time (Principles 06, 11, 13).