Index Us.
Follow via RSS
Menu

The reading room

All articles

Reporting, analysis and practical techniques for making sense of a fast-moving AI industry.

45 articles

Latest first
Analysis7 min read

Computer-use agents need whole-system speed tests

A new preprint shows why model, harness, reasoning and desktop timing must be evaluated together before a computer-use agent is deployed.

Analysis10 min read

Gemini 4 Argon leads one independent suite, but access is still restricted

Gemini 4 Argon has a strong independent benchmark result, but most teams cannot test it yet. Treat the score as a trial prompt, not a deployment decision.

Analysis9 min read

NVIDIA OpenShell ships now, while Sentry remains a reference design

NVIDIA has paired a released agent sandbox with a BlueField watchdog design. Teams can test OpenShell now, but Sentry still needs product evidence.

Analysis9 min read

Anthropic's ART result makes data access part of agent evaluation

Claude agents found a new bacteriophage sequence pattern, but ten reruns missed it. Scientific-agent evaluations need evidence-access and reproducibility checks.

Analysis10 min read

Gemini 3.8 TTS makes voice replication a storage and consent decision

Google's new TTS models add voice replication with a consent check, but teams still need to choose retention, revocation, provenance and migration controls.

Analysis10 min read

Google's private AI memory design keeps device-held keys, with two audit gaps

Google designed persistent cloud memory around device-held keys and attested enclaves. A commissioned audit leaves deletion and response-hash limits to verify.

Analysis11 min read

Claude Opus 5.5 cuts token prices, but its API migration changes agent behaviour

Claude Opus 5.5 is cheaper per token, but breaking request, thinking, tool and fallback changes make it a tested migration rather than a model-ID swap.

Analysis11 min read

Grok 4.7 keeps its token price, but task cost still needs measuring

SpaceXAI kept Grok 4.7's token rates unchanged. Independent tests show where it improved — and why output use, harness and caching still decide task cost.

Analysis10 min read

Googlebook makes screen context a laptop procurement question

Google's new laptop platform puts Gemini into the cursor and desktop. Buyers still need a clear map of what stays local and what reaches the cloud.

Analysis8 min read

Qwen-Image-2.1 is downloadable, but its licence does not cover commercial production

Qwen's new image model is ready to evaluate locally, but the published research licence requires separate terms for commercial use.

Analysis7 min read

California's AI order speeds up auditor rules, not a kill-switch mandate

California has accelerated two AI-auditor programmes and ordered recommendations on frontier-model shutdown controls. It has not yet imposed a kill-switch rule.

Analysis11 min read

OpenAI’s Australian youth blueprint links product controls to policy

OpenAI’s Australian youth blueprint links deployed ChatGPT teen controls to proposals for age assurance, crisis response and independent assessment.

Analysis9 min read

File-review benchmark finds frontier agents overclaimed coverage

A preprint found frequent gaps between agents’ file-review coverage and their final reports. Teams need trace-based evidence before accepting completion claims.

Analysis11 min read

Red-team study finds coding-agent approval monitors need wider context

A preprint reports high bypass rates under a hostile-agent setup. Its useful lesson is to test monitor coverage, context boundaries and containment together.

Analysis10 min read

Google CC turns household coordination into a shared permission problem

Google’s experimental CC agent combines selected household information, memory and actions. The test is whether permission holds through sharing, retention and revocation.

Analysis9 min read

OpenAI Sponsored Agents turn an ad click into a commercial conversation

OpenAI is testing advertiser-sponsored agents in ChatGPT. Operators should test persistent disclosure, claim accuracy, data boundaries and safe completion.

Analysis10 min read

An AI patching benchmark dispute exposes an evidence gap

Trail of Bits disputes 1Password's 26% clean-fix headline. The useful lesson is to measure exploit blocking, complete repair, regressions and review separately.

Analysis9 min read

Cloudflare separates AI training refusal from search crawler blocking

Cloudflare's Disallow AI Training setting keeps designated mixed-use crawlers accessible for search while publishing a training preference. Bing and AI summaries remain separate.

Analysis10 min read

Microsoft AI's draft code sets agent limits before evidence of enforcement

Microsoft AI has proposed rules for future MAI models. Buyers should separate the planned control hierarchy from tested behaviour in a specific deployment.

Analysis9 min read

ZGCM-1 releases a 7B model and training recipe, with conditions on data access

ZGCM-1 publishes weights, code and a large training-data collection. Its open-release claim needs a closer reading of the dataset gate, upstream rights and unreplicated results.

Analysis8 min read

A private-suite coding study finds no average native-harness winner

A same-model agent study leaves the average harness winner unresolved, while workload, completion behaviour and corrected cost telemetry change the practical decision.

Analysis10 min read

Agentforce’s long-horizon runtime changes what operators must test

Salesforce says Hunter can pursue goals for weeks, but remains in pilot. Buyers should test recovery, approvals, memory and stopping conditions before its planned November release.

Analysis11 min read

Anthropic commits to embedded evaluators as OpenAI says it will follow

Anthropic has promised outside evaluators ongoing internal access and publication rights. The test is whether access, independence and redactions become inspectable practice.

Analysis11 min read

OpenAI’s Habitat shows why bounded work matters at scale

OpenAI’s storage engineering account shows how constrained queries, scheduler visibility and connection-pool feedback can matter more than a headline throughput figure.

Analysis8 min read

terms.txt proposes a negotiation layer for agentic web access

A new preprint proposes a web-access exchange for identity, purpose, delegation, payment and receipts. The boundary is useful; interoperability and production evidence are limited.

Analysis9 min read

A2ABreak maps security gaps beyond A2A’s mandatory authorisation

A new protocol analysis argues that authenticated A2A calls still need explicit context ownership, delegation lineage and capability assurance.

Analysis10 min read

OpenAI’s Agents API moves the harness into the platform

OpenAI now manages agent sessions, compaction and recovery, while builders still own tools, environments, authority, data policy and outcome checks.

Analysis9 min read

GPT-Live-1 makes interruption a systems test

OpenAI’s full-duplex voice API can keep talking while an agent works, but interrupting speech does not automatically cancel backend work. Evaluate both control loops.

Analysis9 min read

Subagents outperform inline skills only when the hand-off is clear

A new agent study finds that isolated subagents can beat inline skills under context pressure, but only when procedures have explicit input and output contracts.

Analysis9 min read

OpenAI’s quantum-lab agent makes the control loop the test

An MIT team used GPT-5.6 Sol to calibrate a six-qubit chip. The useful deployment lesson is to evaluate skills, intervention and stop controls together.

Tools9 min read

ChatGPT Images 2.5 makes repeatable editing the claim to test

OpenAI's new image models promise faster, more controlled edits. Teams should measure preservation, accepted-asset cost and provenance before switching.

Analysis12 min read

Meta Muse's local Mac flaw changes its permission-scope test

A researcher says local code can redirect Muse dictation and expose authentication material. Pause Mac expansion until Meta publishes a verified fix.

Analysis12 min read

Cisco’s agent-risk taxonomy sharpens the security boundary beyond the model

Cisco now separates externally hijacked goals from an agent’s own autonomy failures, changing how teams investigate incidents and test controls.

Tools8 min read

Arm AI Portal puts hardware fit beside model choice

Arm's new AI Portal links optimised models to target hardware, runtimes and deployment paths. Use it to shortlist, then test the complete system.

Analysis8 min read

Harbor shows why agent benchmark scores need a version and harness

Harbor unifies dozens of agent benchmarks, but its live task set and judge have already changed. Treat every score as a versioned system result.

Analysis12 min read

OpenAI's Australian account changes the agent incident-response test

OpenAI has detailed four Australian government site interactions, its mid-August discovery and staggered notices. The account corrects the record without verifying containment.

News8 min read

GitHub HydraFusion turns model choice into a runtime decision

GitHub's HydraFusion preview routes coding tasks through single, cascade or critique workflows. Its cost claims merit a bounded, task-level trial.

News7 min read

AI news this week: five changes from 24–30 August 2026

A source-checked account of the AI releases and incidents from 24–30 August 2026 that carry practical consequences beyond launch-day headlines.

Tools7 min read

Cloudflare Kitesurf review: a lean browser for bounded agent work

A documentation and hands-on review of Cloudflare Kitesurf, including where its lightweight browser engine works, where Chromium remains safer and how to test it.

Guides8 min read

How to verify an AI tool announcement before you trust it

A practical seven-step method for separating a useful AI release from a polished demo, incomplete rollout or recycled capability.

Techniques8 min read

Build a useful AI evaluation from twenty real cases

A practical method for creating a compact AI evaluation that measures quality, review time, cost and failure severity on work your organisation recognises.

Analysis7 min read

MCP explained without the protocol soup

A practical explanation of what the Model Context Protocol connects, where permissions sit, and what teams still need to secure themselves.

Analysis7 min read

What an AI benchmark can and cannot tell you

Learn how to read AI benchmark scores by checking the task, test conditions, contamination risk, grader and fit with the work you actually need done.

Guides7 min read

How to choose an AI model for the job

A vendor-neutral method for selecting an AI model by task quality, privacy, tool support, latency, cost and the failures your workflow can tolerate.

Analysis9 min read

How to read AI industry news without getting lost in the hype

A practical framework for reading fast-moving AI coverage: identify the change, inspect the evidence, find the affected user and choose a proportionate next step.