Reasoning is leaving the screen: Astra folds it into latent space, Gemini spreads it over cheap tokens, and a BCI paper prices the cost — 28.4% of high-confidence outputs are unfaithful

OpenAI says Astra can locate and exploit unknown vulnerabilities in computer systems without human guidance, scoring a perfect result on ExploitBench — an evaluation of LLM hacking ability — and surfacing two previously unknown zero-days in OpenAI's modified version of the test. The company says it plans to make Astra available "soon" while restricting its most advanced cybersecurity capabilities at first, paired with improved abuse detection, jailbreak prevention, risk-tiered account restrictions and chain-of-thought monitoring (TechCrunch, 2026-09-01). The Astra family was first announced on August 1 and updated on August 6 with solutions to ten long-unsolved mathematical problems, all formalised in Lean at roughly $2,000 of API tokens (The Decoder, 2026-08).

Hospitals are among the most poorly defended large organisations on the network, and a model that finds zero-days on its own is a tool for defenders and attackers alike. More pointed: OpenAI itself lists chain-of-thought monitoring among its safeguards — and if the recurrent-depth reports are accurate, there is less chain of thought left to monitor. TechCrunch also notes the safety claims currently lack independent third-party verification.

"Recurrent depth" so far rests on community readings and secondary coverage; OpenAI has published no checkable technical report. The perfect ExploitBench score and the two zero-days are vendor-reported and not independently audited. Sebastian Raschka's September analysis pushes back on the idea that looping necessarily suppresses visible reasoning, arguing that reusing layers does not by itself hide a chain of thought (Raschka, 2026).

TechCrunch, 2026-09-01 ／ The Decoder ／ Sebastian Raschka

Gemini 3.8 Flash scores 59 on the Artificial Analysis Intelligence Index at high reasoning — three points above 3.7 Flash and level with GPT-5.6 Sol and Grok 4.6 — with a 1M-token context window, text/image/video/speech input and roughly 300 output tokens per second. Pricing is discounted to $0.75/$3.75 per million input/output tokens through the end of 2026, with a standard rate of $1.50/$7.50 and about $0.58 per task. It is Google's fourth Flash model in under four months (Artificial Analysis, 2026-09-02). The Flash Cyber variant, launched the same day for security researchers finding vulnerabilities and writing patches, beat Claude Opus 5 and GPT-5.6 Sol on 9 of 16 tests, scoring 73.7% on DeepSWE-1.1 and 86.2% on CyberGym C/C++ vulnerability detection (SiliconANGLE, 2026-09-02).

For health systems, what decides whether AI reaches the whole hospital is usually unit cost rather than peak capability: at the discounted rate, sweeping a 200,000-token inpatient chart costs roughly $0.15 in input tokens. Note also the governance split — Google turned the security capability into a public, purchasable variant while OpenAI restricted access to its own. Google's Tulsee Doshi and Raluca Ada Popa say the model "exhibits greater diligence — executing extra reasoning steps, and calling tools iteratively" on complex tasks.

The benchmark figures come from Google's own reporting and a third-party index, and none of them are medical tasks; no clinical metrics such as MedQA or image interpretation accompany this release. The discount expires at the end of 2026, so long-run budgeting should assume the $1.50/$7.50 standard rate.

Artificial Analysis, 2026-09-02 ／ SiliconANGLE, 2026-09-02

A team drawn largely from radiology and informatics (Gorenshtein, Omar, Klang, Barash and colleagues) posted "Intent Drift in LLM-Assisted BCI Communication" on medRxiv, simulating P300-speller decoder corruption to test how faithfully 20 open-weight models repair text for brain–computer interface users. Drift climbed from 2.2% with no decoder error to 60.3% at high corruption; 28.4% of the models' own high-confidence outputs were unfaithful to the intended message; discrimination was moderate (AUROC 0.83) but calibration was poor (expected calibration error 0.32); the best correction policy still produced wrong messages in 18% of cases; clinically critical messages drifted slightly more, not less; and physician review agreed with automated scoring only moderately (kappa 0.41) (medRxiv, 2026-08-27).

This puts a number on model confabulation in the one setting where it is least tolerable: the user is a patient who has lost the ability to speak, and the output is taken as their own words. The authors' phrase is "fluent, highly confident lies" — the failure is not an uncorrected typo but a changed intent, asserted with conviction. An ECE of 0.32 means the confidence score is close to useless as a gate, which invalidates an entire class of deployment designs built on confidence thresholds.

This is an in-silico benchmark: decoder errors were simulated rather than drawn from real BCI users, and it is a preprint that has not been peer reviewed. All 20 models are open-weight, with frontier closed models (Astra, Gemini 3.8, Claude Opus 5) excluded, so the figures cannot be extrapolated directly to the commercial models actually deployed in clinics.

medRxiv 10.64898/2026.08.20.26360939, 2026-08-27

An LLM agent called ScreenAgent was built to automate the first-pass screening stage of medical meta-analyses. On the reported figures it processed 201,064 records for a total of $855.91 — about 0.43 cents each — with an internal sensitivity of 97.7% (43 of 44 eligible studies found) and a 99.4% workload reduction for human reviewers. Agent-to-human agreement reached a Cohen kappa of 0.75, above the 0.64 measured between human reviewers, with external validation sensitivities of 95.9% and 97.4%. The authors use a hybrid design in which a second LLM verifies results and human experts retain the final call (Yesil Science, 2026-08-30).

Systematic review is the cost bottleneck of evidence-based medicine; first-pass screening for a large meta-analysis typically occupies two researchers for months. If the 0.43-cents-per-record figure holds, the refresh cycle for clinical guidelines stops being limited by labour and starts being limited by how fast primary research appears. The contrast with the previous item is instructive: the same class of technology is badly calibrated when speaking on a patient's behalf, yet robust on a task with an explicit recall target and a second pass to catch its errors. Task structure, not raw model strength, decides whether an agent is usable.

We could only reach secondary coverage for this item and could not verify the numbers against the medRxiv original, so every figure above rests on that report. The system missed one eligible study in testing — and in a meta-analysis, the missed study can be the one that changes the conclusion. A 97.7% sensitivity sounds high, but evidence synthesis conventionally demands a stricter bar.

Yesil Science, 2026-08-30 ／ The Health AI Brief, week of 2026-08-31

GE HealthCare's Photonova Spectra received CE marking on August 31. Where conventional CT detectors first convert X-rays into visible light, the Deep Silicon detector measures individual X-ray photons and their energies directly, so every scan yields high-definition images and spectral data at once, across neurology, oncology, musculoskeletal, thoracic and cardiac imaging. The system already held FDA 510(k) clearance and Japanese approval as of March 2026. Chad Rowland, executive director of global CT at GE HealthCare, said: "At its core is our Deep Silicon detector technology, which enables Photonova Spectra to capture remarkably detailed, spectral information in every scan." (MobiHealthNews, 2026-08-31)

In a week dominated by software, this is one of the few breakthroughs at the physics layer — and it determines what the models above it get to see. Spectral CT carries material composition in every voxel rather than a single HU value, which effectively swaps imaging AI's input from a greyscale picture to a multi-channel measurement. Imaging foundation models trained on conventional CT cannot simply ingest that data, and the cost of retraining and revalidation lands on the hospitals that buy photon-counting scanners.

The coverage provides no figures for resolution, dose reduction or clinical outcomes, and no head-to-head comparison against existing photon-counting platforms such as Siemens' Naeotom. CE marking is a licence to sell, not evidence of clinical benefit.

MobiHealthNews, 2026-08-31

Stryker announced the first clinical use of SportSuite Vision on September 1. It is a video see-through augmented reality head-mounted display that folds arthroscopic video, the HipCheck software, HipMap femoroacetabular impingement analysis and CT imaging into a customisable spatial computing environment, overlaying digital clinical information inside the surgeon's direct field of view. The FDA granted De Novo authorisation on July 17, 2026, the first approval for intraoperative use with Apple Vision Pro; the inaugural case was performed by Duke Health orthopaedic surgeon Chad Mather III, with the indication limited to hip arthroscopy for FAI and labral repair (Stryker, 2026-09-01; STAT, 2026-09-01).

De Novo rather than 510(k) means the FDA found no comparable predicate, so this creates a new classification and special controls for intraoperative head-mounted displays — later entrants can follow via 510(k). Dr Mather's account points to ergonomics rather than immersion as the real value: "spatial computing enabled me to customize the placement of key clinical information to fit my workflow and access it within the sterile field." This is the first time consumer hardware has entered the sterile field as the primary display.

Stryker's announcement mentions no AI component — this is a display and integration system, not automated interpretation. The surgeon on the first case is also a Stryker consultant, and with a single case there are no comparative data on operative time, complications or learning curve. Most of the STAT piece sits behind a paywall, so this item rests chiefly on Stryker's own press release.

Stryker Newsroom, 2026-09-01 ／ STAT News, 2026-09-01

Medtronic announced on September 1 an investment of roughly $700 million for distribution rights to Hong Kong-based Cornerstone Robotics' Sentire surgical system in select approved markets outside the United States. Sentire was developed entirely in-house, features dual-console capability and an immersive console design, received CE marking in May 2026 for minimally invasive general, gynaecologic, thoracic and urologic procedures, and holds market approval in China and Singapore. Medtronic's own Hugo RAS operates in more than 35 countries across six continents, with global procedures projected to exceed 50,000 by fiscal year end, and a digital ecosystem that includes real-time AI and the Touch Surgery platform (Medtronic, 2026-09-01).

Value in surgical robotics is migrating from the arm to the data it generates: every installed system is a continuous source of operative video and instrument kinematics, which is the one training corpus for surgical AI that cannot easily be substituted. By running Hugo and Sentire together, Medtronic trades distribution for data scale, in contrast to Intuitive Surgical's single-platform strategy — making it the second player outside the da Vinci franchise with a plausible path to pooling data across machine types. Matt Anderson, senior vice president of Medtronic's surgical division, said "surgeons, health systems and markets have varying needs and preferences," calling for flexibility across clinical settings.

The release does not say whether data generated by Sentire feeds Medtronic's digital ecosystem, nor how the $700 million splits between equity and distribution rights, nor Sentire's installed base or procedure volume. The 50,000 figure is a Hugo projection, not a realised number.

Medtronic Newsroom, 2026-09-01 ／ MobiHealthNews, 2026-09-01