The boundaries around agents and generated work

GPT-6 Cyber tests exploit skill and authorized scope, Agent Substrate separates session state from compute, and Gemini Omni 1.2 Flash changes short-video production.

3 stories · Yesterday in AI

TL;DR

The short version
  1. GPT-6 Cyber improves OpenAI’s exploit and scope-following scores, while initial access remains restricted and organizations retain responsibility for authorized boundaries.
  2. Google’s Agent Substrate pools compute while preserving agent state; evaluate recovery and isolation alongside its reported fast resumes.
  3. Gemini Omni 1.2 Flash adds 15-second clips and conversational edits at a lower listed rate; judge savings by approved outputs and review effort.

Enterprise Ops

1 story

Google separates agent state from compute in Agent Substrate

Google’s Agent Substrate separates saved agent sessions from pooled compute; its CNCF announcement and fast-resume claims create a practical restore test.

As a leader: Compare saved idle compute with storage, restoration and operational costs using a representative agent workload.

As a practitioner: Pause a session with uncommitted work, restore it on another worker, and verify both state and access boundaries.

Google Open Source Blog ↗Vendor documentation

Vendors

1 story

Gemini Omni 1.2 Flash expands clips and lowers generation prices

Google announces 15-second video and conversational editing in Gemini Omni 1.2 Flash, with a lower listed rate and remaining motion limitations.

As a leader: Budget a pilot around approved clips and review effort, using representative product references and acceptance criteria.

As a practitioner: Check reference details and motion after each conversational edit, and record retries before comparing production costs.

Google ↗Vendor documentation

Security

1 story

GPT-6 Cyber pairs stronger exploit results with restricted access

OpenAI reports stronger exploit and scope-following results for GPT-6 Cyber, with initial access limited to selected approved Daybreak Red customers.

As a leader: Treat access approval, authorized targets and isolation as separate readiness decisions before commissioning a pilot.

As a practitioner: Record allowed targets and stop conditions, then test scope compliance separately from successful vulnerability exploitation.

OpenAI ↗Vendor documentation