All the research, one tab.
Every research item behind the daily briefs, newest first, in fast pages. Filter by topic, open any card's source, share any page — the URL is the state. The live map stays on the Brief page.
- Research
Anthropic says Opus 5 is its hardest model yet to prompt-inject — the evidence sits in the system card
Prompt injection is still the open hole in every tool-using agent. A vendor claim of improvement is worth tracking, but until it is independently measured your permission gates stay exactly as load-bearing as they were.
- Research
One robot model, roughly 9,000 different hands: Generalist’s GEN-1 transfers across end effectors
Retraining per gripper is the tax that keeps robot deployments bespoke; a model that transfers across hardware is what turns a pilot into a fleet.
- Research
A team that ran fully autonomous coding agents for a year reports the catch: codebases decay
First-hand, long-duration evidence for keeping deterministic gates and review layers in agent-driven development rather than trusting full autonomy.
- Research✓ verified
The OpenAI–Hugging Face breach traces back to a human mistake: a sandbox left open to the internet
Teams running agentic evals with relaxed refusals should audit sandbox egress first — the failure mode is ordinary infrastructure, not exotic AI.
- Research
McKinsey puts a number on AI in construction: 39% of nonphysical work is automatable
A rare sector-specific automation estimate with a phasing timeline — useful whether you are deciding what to pilot first or what not to build in-house at all.
- Research✓ verified
Meta's open vision models cut a month of scientific image labeling to 15 minutes at Berkeley Lab
A real, measured deployment — not a benchmark claim — showing open vision models replacing weeks of expert labeling with minutes, a concrete reference for any team weighing open models against paid annotation.
- Research✓ verified
Loops, graphs, and anchors: how to orchestrate recurring agent work
If you run recurring agent jobs — nightly content builds, generated artifacts, scheduled reports — the practical lesson is to match the architecture to the job's real judgment need, not to reach for a framework. Most recurring work needs a fixed pipeline with one narrow place where an agent authors something, and the failure almost never lives in the loop-versus-graph choice — it lives in the seam: cleanup and 'did every output actually save' steps that were written as an instruction in the agent's prompt instead of as a hard step in the surrounding script.
- Research✓ verified
How AI Uni's agents remember: two kinds of memory, and why the durable one is layered
If you build agents that must remember anything across a killed session, a fresh subagent, or a model swap, the reusable lesson is that no single memory file can do the job: each layer here exists because a different, specific way work got lost was actually observed, and each delivers the fact at a different moment to a different audience. The honest-limits section is the most useful part — writing something down is not the same as guaranteeing it reaches the right agent later, and this design says so out loud.
- Research✓ verified
Think it, and the robot does it: BrainCo shows a brain-to-robot platform that reacts in under 200 milliseconds
Brain-computer interfaces have mostly lived in research demos and rehab clinics; a sub-200ms decode loop that plugs into off-the-shelf robots is what turns "thought control" from a lab trick into a control scheme other robotics teams can actually build on.
- Research✓ verified
AI systems are out-counterexampling human mathematicians — with machine-checked proofs
Generate-then-machine-verify turns AI output from a claim into a checkable artifact — a pattern that generalizes directly to code.
- Research
Google DeepMind and Isomorphic Labs publish their playbook for keeping AI biology models safe — and useful
Biology is where AI capability and AI risk get most concrete — this is what the leading lab's actual mitigation posture looks like.
- Research
NVIDIA Isaac Lab-Arena: an open framework to evaluate general-purpose robot policies
A robot policy that works in a demo can fail in the world — shared, large-scale evaluation in simulation is the missing benchmark layer physical AI needs before it is trusted off the lab floor.
- Research✓ verified
Making an autonomous work loop survive the seams: how an agent loop was designed to resume its own goal from disk after a killed session (built + reviewed, dry-run pending)
If you're building an agent that has to keep working across session death, context compaction, or a model downgrade, the hard part isn't retrieving state — it's proving the loop resumes the RIGHT state and can't run away, ship on its own, or grant itself a fresh budget every restart. This is a worked, honestly-graded design for exactly that: a durable state cell, a cumulative budget that survives restarts, a halt-and-hand-back rule instead of a silent spin, and a never-self-ship gate — with the parts that are green-in-tests kept clearly separate from the parts still pending a live dry-run.
- Research✓ verified
Cross-session agent memory in 2026: the platform now ships natively what most teams still hand-roll
If you're building an agent that needs to remember anything across sessions, check whether Anthropic's native memory tool already does what you were about to hand-roll — and if you're already building one, the field's clearest fix for the write-back/consolidation gap is a scheduled, importance-triggered pass, not more discipline.
- Research
You can't reliably detect AI-written code — the detectors aren't trustworthy enough to stand alone
Don't hang a pass/fail — or an accusation — on an AI-code detector. Treat its output as a hint to look closer, never as the verdict.
- Research
The one genuinely new 2026 tool: a cryptographic 'receipt' that proves a tool actually ran
There's finally a cheap way to prove a tool ran and what it returned — but only when you run the tool. A receipt a user hands you for a tool they ran themselves proves nothing.
- Research
In 2026, the most AI-resistant way to check understanding is still a live human conversation
No tool can tell you whether someone actually understands their work. In 2026, the reliable signal is still the oldest one: ask them to explain it, live.
- Research✓ verified
Can you prove an AI agent actually did the work? What today's tools can and can't verify
If your team accepts work an AI helped produce, this is the honest map of what's actually checkable in 2026 — and where a human still has to be the judge.
- Research
The tools that flag 'this looks AI-written' are biased against non-native English writers
Automated 'is this AI?' signals don't just misfire — they misfire unevenly, penalizing people who write in a second language. That's a fairness problem, not just a noise problem.
- Research
'My AI read your repo and it checks out' is a claim, not proof
Anything an agent reports about work done on a machine you don't control is unverified by default. If it matters, re-run the check yourself instead of trusting the summary.
- Research
Anthropic + AE Studio publish "modular pretraining" for gating dual-use model capabilities (GRAM)
Early research, not a shipping feature: it points toward a future where specific model capabilities can be switched off without retraining, but it is a lab result explicitly not in any production Claude today — nothing to adopt yet.
- Research✓ verified
Anthropic + Glasswing partners propose an industry jailbreak-severity score
A common severity scale would change when a lab (or a regulator) decides a model must be pulled — directly relevant to how agent products get governed.
- Research
Microsoft Research: agent skills as trainable parameters (SkillOpt) + token-efficient agent memory (Memora)
Concrete, buildable levers: systematically optimizing your skill/instruction files (rather than the model) and compressing agent memory can raise quality and cut token cost — directly applicable if you author skills or run long-horizon agents. Figures are the authors' reported results; validate before quoting.
- Research
Google DeepMind launches a Robotics Accelerator, putting Gemini robotics models in startups' hands
When a frontier lab puts its vision-language-action models directly in startups' hands, embodied AI stops being a lab demo and starts becoming an ecosystem — the same pattern that scaled language-model apps.