Horizon Lens — 11 November 2026
A machine-checked proof for the code behind Signal messages
Google DeepMind says its researchers proved that the compiled code behind Signal’s core cryptographic protocol meets a set of security requirements. A protocol describes how a conversation should stay private. The research went a step further and checked whether the code that actually runs on devices delivers the protections the researchers specified.
Human specialists reviewed each mathematical specification. Agents then wrote proofs in Lean, a language with a checker that verifies the steps. DeepMind says the process took four weeks. The protections examined included keeping messages secure on either side of a period when a device was temporarily compromised.
Analysis
The important human decision came before the proof. Choosing the property to check determines what readers can conclude afterwards. Someone trying to assess the privacy of an entire messaging app needs to know exactly which part of the software each guarantee covers.
Action
When you see a claim about verified software, look for three things together. Which property did the researchers specify, which implementation did they check, and can someone else inspect the proof? DeepMind says it is making its verification materials public.
Seagate’s new hybrid drive comes with a shopping list
Seagate’s FireCuda Forge puts a hard drive and an SSD together in one enclosure. There is one detail worth catching before comparing prices. Ars Technica reports that the $699 model includes an 8 TB hard drive and the $899 model includes a 20 TB hard drive. Both require an SSD that the customer supplies separately.
According to Ars, the enclosure accepts M.2 2280 NVMe drives using PCIe Gen3 or later, up to 1 TB. Seagate says its software learns which files are in demand and keeps them quick to access. Ars found the announcement left open what happens to the combined volume if one drive fails.
Analysis
Imagine editing the same few projects repeatedly, then opening an old archive. Those are two different tests of an enclosure that learns which files to keep handy. A transfer speed quoted for the system tells you little about which files have made it onto the fast drive.
Action
Price the enclosure with a compatible SSD and check the required read and write speeds before reusing a spare. For a working archive, ask what can be recovered if either drive fails before counting on the combined volume for storage.
What businesses can see when ChatGPT mentions them
OpenAI has opened a beta that lets eligible businesses claim their profiles in ChatGPT. After verification, they can correct company details and supply product feeds with prices and availability. The company says eligible businesses can claim a profile wherever ChatGPT is available. Shopify merchants can connect their stores to help with verification.
The dashboard can show mentions, website clicks and common topics in conversations that cite the business’s site. OpenAI says it applies minimum activity thresholds and keeps conversations private. A separate feature lets customers choose to talk to a business’s own agent. That initially covers eligible US ecommerce and online service businesses.
Analysis
For a shop owner, discovery and conversion are different questions. A mention tells you the shop came up. A website click records another step. Neither figure alone tells you whether someone found the particular item they wanted or bought it.
Action
Start with a product whose price or stock changes often. Check what the verified feed says against what a shopper sees on the shop’s own site. When interpreting an empty dashboard, remember that OpenAI requires a minimum amount of activity before it shows insights.
A small assistant loses ground as instructions pile up
Tether AI Research has introduced BotPsy, a model with 800 million parameters designed to run on a device and call tools such as a calendar or email app. Its authors position it for routine tasks of low to medium difficulty. They say it can make mistakes and advise against using it for high-stakes work.
On one test that tracks instructions across three turns in eight languages, the authors report a score of 84.3 on the first turn and 51.0 on the third. It began ahead of the other four models in their comparison. Two competing models finished the third turn above BotPsy.
Analysis
A single calendar request is a different assignment from a conversation in which you change the time, narrow the guest list and expect the earlier constraints to carry over. The gap between those two scores gives a developer a concrete reason to test the whole exchange.
Action
Try that sequence with a disposable calendar. Give the assistant one instruction at a time, then inspect its proposed event after the third turn. Check the details you carried forward as carefully as the change you asked for most recently.
Gemini Carbon reaches paying developers with a different cost question
Google has expanded access to Gemini 4 Carbon following an earlier release to a group of cybersecurity defenders. Paying developers and enterprises now have API access. Google also says the model is rolling out to Google AI Ultra subscribers in the Gemini app, Antigravity and AI Mode in Search.
Through November 30, Carbon matches Argon at $2 per million input tokens and $10 per million output tokens. Google says public Carbon will then cost $4 and $20 respectively. It says Carbon uses more tokens for difficult tasks and is previewing a way to coordinate several agents in Antigravity.
Analysis
During the introductory period, equal token prices leave room for different bills. For a developer commissioning a report, the important quantity is what it costs to get a report that survives inspection. Longer reasoning can be worthwhile if the answer clears a check that the cheaper attempt failed.
Action
Choose one difficult task with an answer you can check. Run it with both models, retain the outputs, and record the tokens consumed alongside whether each answer passed. That gives your team a comparison tied to the work it actually does.