AI and machine learning has stopped being a field and become an instrument layer: the same statistical machinery is now the default tool for reading biosensors, solving crystal structures, tuning epitaxy, writing hardware description code and driving radio beams, while the agentic layer built on top of large language models is measurably insecure.
Summary
“AI and machine learning” as a technology entry is not a single artefact but a method stack: supervised and unsupervised models over instrument data, generative models (diffusion, autoregressive language models) used as learned priors or code writers, and reinforcement learning used as a controller. What decides whether it works in any given place is not the algorithm but three parameters: whether a labelled or simulatable dataset exists at sufficient scale, whether the target quantity is actually recoverable from the measured signal, and whether errors are cheap to detect downstream.
The evidence slice here shows all three patterns at once. As a decoder, a twelve-chirality carbon nanotube sensor array plus a convolutional autoencoder reached 84.6% cross-validated accuracy (AUC 0.87) separating ALS serum from controls in a 39-person cohort, and unsupervised learning on in-situ RHEED patterns separated high- from low-quality GaSe epitaxial films, correlating strongly with rocking-curve FWHM but weakly with surface roughness. As a generative prior, a diffusion model recovers crystal structures from powder X-ray diffraction patterns, a phase-loss inverse problem, with polymorph families held out together to prevent memorisation. As a surrogate, language-model representations estimate CPU module power directly from source-level design information without running simulation.
The second, harder half of the picture is the agentic layer. Language models are now being pushed into roles where their output is executed rather than read: generating SystemVerilog assertions for formal verification, generating VHDL, driving browsers, loading runtime “skills”. Here the measured numbers are poor. Twelve recent models produced assertion sets that compile at the module level 67.1% of the time and, although 82.1% of generated assertions prove non-vacuously, they detect only 70.2% of eligible injected faults and cover 36.2% of the formal core. Safety layers fare worse: existing skill scanners detect 2-8% (open source) and 9-17% (commercial) of cross-modal language-and-code attacks on agent skills, agents that refuse monolithic harmful tasks comply once the task is decomposed into benign subtasks, and agentic browsers routinely break the same-origin policy that has anchored web security for decades.
Read together: ML as a measurement and design instrument is delivering; ML as an autonomous actor is at the stage where the attack papers are outrunning the defence papers.
Evidence base
- 9 June 2026: a twelve-chirality DNA-functionalised carbon nanotube sensor array with a dual-objective convolutional autoencoder reached 84.6% cross-validated accuracy and AUC 0.87 on serum from 20 ALS patients and 19 matched controls.
- 15 June 2026: across twelve recent LLMs, generated SystemVerilog assertions gave a 67.1% module-level compile rate, 82.1% non-vacuous proofs, but detected only 70.2% of eligible injected faults and covered 36.2% of the formal core.
- 15 June 2026: open-source and commercial agent-skill scanners detected only 2-8% and 9-17% respectively of language-and-code cross-modal attacks across 13 attack categories.
- 15 June 2026: a two-dimensional programmable waveguide with about 10^4 programmable spatial degrees of freedom performed neural network inference on up to 49-dimensional vectors in a single pass with no digital pre- or post-processing.
- 15 June 2026: guardrail denial-of-service payloads optimised on a single open-source surrogate transferred successfully to eight leading model backbones.
- 15 June 2026: merged amplitude encoding cut quantum circuit executions per forward pass by a factor of n for 1-2 extra qubits, with no significant trainability difference from the baseline in 28 of 30 comparisons (p > 0.05).
- 15 June 2026: a cross-sector perspective concluded ML accelerates materials exploration only when supported by well-curated datasets with clear provenance and uncertainty quantification.
Open questions
- Does the ALS protein-corona classifier hold up in an independent, larger cohort, and does AUC 0.87 on n=39 survive external validation and pre-symptomatic samples?
- Can install-time detection of cross-modal agent skill attacks be pushed from the current 2-17% range to a rate that would satisfy an enterprise security review, and does that defence generalise beyond the 13 attack categories it was tuned on?
- Do LLM-generated verification artefacts reach the point where formal core coverage, currently 36.2%, is high enough to reduce verification headcount rather than merely reallocate it?
- Will experimental facilities actually adopt the dataset provenance and uncertainty-quantification discipline that the materials community says ML acceleration depends on, or does the data bottleneck persist?
Assessment drafted 2026-08-31 from up to 18 KB sources using the technology-scorecard framework; scores are a draft read pending review.