AI developments, by field

BioBasicsBench tests whether agents respect experimental evidence

Across 100 routine wet-lab tasks, OpenAI reports 66.7% mean reward for GPT-6 Astra at xhigh, versus 43.8% for GPT-5.6 Sol. These are shares of available rubric credit, not complete-task success rates. The report still finds assumptions overriding evidence.

What to watchbetter average scores do not establish reliable lab judgments; an evaluation must also test when evidence is insufficient.

Kimi Hosted Agents makes the runtime a managed service

Moonshot's 2026-11-05 announcement describes isolated sandboxes, network controls, interruptible sessions and audit exports. Its separate pre-inference security hook is currently offered selectively to enterprise-verified users.

What to watchruntime controls become part of the product, but the announcement does not independently establish isolation or productivity gains. · managed-agents (source)

Harness-S

rescue failed training attempts with sibling harnesses, then return the trained policy to its original harness. Read it for separating training assistance from deployment dependencies; the abstract reports improvements without numerical effect sizes.

Hybrid Computer Use

combine code, accessibility and graphical interaction, with protection against stale-state actions during channel switching. Read it for interface-aware evaluation; reported OSWorld 2.0 gains have no numerical scores in the abstract. First submitted 2026-10-30.

Kimi Hosted Agents moves the agent runtime into a managed service

Moonshot's announcement, dated 2026-11-05, describes hosted execution with separate sandboxes, network permissions, interruptible sessions and exportable audit traces.

What to watchthese controls give teams explicit points to constrain and inspect execution. They are vendor-described capabilities, not independently verified isolation or productivity gains. (primary)

Harness-S

after all initial attempts under a base harness fail, use alternative harnesses to collect replacement training trajectories. The authors report improvements when the trained policy returns to the original harness. Why read it: it separates training assistance from deployment dependencies; the abstract provides no numerical effect sizes. Version history: v1, 2026-11-03 04:03:44 UTC. (primary v1)

Hybrid Computer Use

combine code, accessibility and graphical interaction, with checks against acting on stale state when switching channels. Authors report better task success and efficiency across the evaluated models. Why read it: interface choices belong in agent comparisons; the abstract supplies no numerical scores or latency measurements. Version history: v1, 2026-10-30 08:38:29 UTC. (primary v1)

GPT-6 Cyber pairs stronger security specialization with scope training

OpenAI's new model requires Daybreak Red approval. In its controlled honeypot evaluation, out-of-scope exploitation falls from 5.51% to 0.07% without deployment safety instructions; the observed rate with those instructions is 0%.

What to watchtask boundaries are measurable, but the system card also records weaker final blocker reporting than Astra. These results do not guarantee safe deployment. (access)