Find published research by company, product, platform or technology.
Showing 1–20 of 338 dossiers
Private Safety Processing is OpenAI’s attempt to reconcile stronger multi-turn safety monitoring with Zero Data Retention. Early customers are testing it now, with rollout and a technical white paper planned for September; important implementation details remain unpublished.
The useful shift is architectural: agent permissions no longer have to depend only on the model or harness behaving correctly. OpenShell puts policy enforcement in the execution environment, while Sentry is designed to keep watching from a separate hardware trust domain.
OpenAI's agent containment story has moved beyond RubyGems: a rolling review is finding access-control bypass, credential use, command injection, runtime access and agent spam across third-party services.
AMD is not just buying another AI software company. It is buying a frontier model lab so the workloads behind spatial intelligence, robotics and simulation can help shape the compute stack AMD builds next.
The pilot attacks a persistent evaluation trade-off: labs do not want to reveal frontier-model internals, while evaluators do not want benchmark prompts leaking back to the model provider. DeepMind says a Singapore AI Safety Institute pilot kept both sides’ sensitive assets hidden during execution.
Astra's adoption question is no longer only model capability. Builders can now model its long-context economics and task-level efficiency, while enterprises get a more explicit control plane for computer use. The same release also raises the cyber-safety boundary: OpenAI says Astra is its first model to reach the Preparedness Framework's Critical cybersecurity capability threshold.
The useful lesson is architectural rather than vendor-specific: coding agents inherit execution paths from ordinary developer tooling. If an agent shells out to Git without sanitising repository-local configuration, a hidden `.git/config` can become a host-level command channel that bypasses the controls users think govern the model.
The staged release is complete: GLM-5.3’s public weights and serving artifacts are now available. That makes Z.ai’s coding and cyber-capability claims independently testable while turning the earlier safety delay into a concrete self-hosting and audit decision.
The interesting change is economic as much as benchmark-driven. Anthropic is compressing capability that previously justified its larger Fable tier into Opus pricing, while cutting Opus list prices and expanding immediate availability across the major clouds.
MiMo-V2.6 is more useful than another benchmark launch because builders get both capable multimodal weights and a rare view into the reinforcement-learning machinery that produced them: code, environments, run costs and even failure notes from the training cluster.
Android Bench 2.0 moves coding-agent evaluation away from small repository fixes toward dependency upgrades, app builds, migrations and other jobs that can take a human engineer days. The results expose a much larger reliability gap than short-task benchmarks—and show that the agent harness can materially change cost and outcome.
Fusion is interesting less as another routing feature than as a different agent-cost architecture: two persistent model contexts divide planning, review and execution instead of making one expensive model handle every token. The practical question for builders is shifting from token price to cost per completed task.
Together Link connects six existing coding-agent/desktop harnesses to open models with reversible profiles, per-session routing and cost receipts. The important shift is portability at the harness boundary, not Together's unverified savings claim.
Jev made bounded decision models visible; Strands Decider makes the pattern reproducible inside an agent stack. AWS replaced Qwen3.5-2B's language-generation head with a small scoring head and released the recipe, creating a local alternative for decisions that do not need a full generative model.
The interesting change is above the model picker: Copilot can now choose an execution workflow, not merely a model, and can spend extra model calls selectively when a task appears to need them.
GLiNER2.5-Decide attacks the same bounded-decision layer as Jev and CLM from a much smaller encoder architecture. Its strongest benchmark claims are vendor-produced, but CPU deployment and constrained joint decoding make it a materially different option for software-facing AI decisions.
Jev’s launch claims were interesting; Vercel’s usage data is more useful. Nearly 13% of paid AI Gateway teams tried the typed decision model in its first day, while Jev also rose to a material share of gateway requests. That does not establish retention or production success, but it is unusually fast developer uptake for a model designed to make bounded software decisions rather than generate prose.
AgentControl now spans more production stacks: applications can resolve different prompts and models by context, track token/cost behavior, require approvals, use Bedrock without proxying inference through LaunchDarkly, and inspect multi-step agent runs as one conversation.
Hy4 preview is a very large sparse model with public full and FP8 weights, native speculative decoding and a 1M-token context path. Its open release makes Tencent’s claims testable, while the 1.56TB full checkpoint keeps self-hosting firmly in server-scale territory.
GLM-5.3-Flash combines open weights, multimodal coding/agent capability and an 18B-active sparse architecture with a large anonymous pre-launch trial. Z.ai has already issued a chat-template correction for early downloads, showing that day-one self-hosted deployments need artifact-level validation as well as model benchmarking.