Key points
- DeepMind — A mathematical proof for cryptography. Google DeepMind researchers used agentic AI to produce mathematical proofs — formal verification — of several key cryptography guarantees of the Signal protocol.
- Meta — Meta calls it its smallest and most efficient model, aimed at the tool calls, coding subtasks, routing, extraction and long-document parsing that agents repeat.
- Blender 5.3 — Blender 5.3 is out with support for Gaussian Splatting, plus extended Geometry Nodes bundle capabilities, so working with splat data and fairly complex visual programming is more direct.
DeepMind says it formally verified key Signal cryptography guarantees, Meta AI shipped Muse Ember 1.0 for the short tasks agents repeat, and IGF jurors are refusing to vote for games built with generative AI, DeepMind formally verified key Signal cryptography guarantees with agentic AI, and Meta AI shipped Muse Ember 1.0 for the short tasks agents repeat.
House House disclosed a small amount of AI use in Big Walk, and IGF jurors are refusing to vote
Over the past few years AI tools have become more common in game development, and whether players should be told has been an open argument. This week that argument broke into the open in the indie scene. House House, the studio behind the well-liked co-op puzzle game Big Walk, admitted it used a very small amount of AI — and said so only after the fact. The dispute had been building behind the scenes for months. Jurors and panelists of the Independent Games Festival Awards discussed it in private groups. On Monday, 19 of the 50-plus jurors published a joint statement: they refuse to vote for any game that uses generative AI in any way. They point to a significant increase in submitted games built with generative AI tools, and name the ecological, economic, psychological and labour harms of those tools as their reason. They say such use “is not yet a foregone conclusion in our industry.”
What actually changed here is the relationship between jurors and developers: a tool can be used quietly, but whether and when to say so now affects other people’s votes and livelihoods. Nineteen signatures do not change the awards’ rules, but they turn a matter of tacit understanding into a public position. What to watch next is whether the awards body introduces a disclosure requirement, and whether a late footnote is treated as a mistake or as normal practice.
DeepMind formally verified key Signal cryptography guarantees with AI
Next, something that asks for more patience: a mathematical proof for cryptography.
Google DeepMind researchers used agentic AI to produce mathematical proofs — formal verification — of several key cryptography guarantees of the Signal protocol. Reports call it a landmark for widely used cryptography and for AI itself. The work took about four weeks and is a proof of concept with two meanings. One is independently re-checking the cryptography that protects a large share of messaging, from the Signal app to Google Messages and WhatsApp. The other is showing what DeepMind’s new AI framework, AlphaProof Code, can do in formal verification. More broadly, AI-assisted formal verification is getting attention, but whether it can be used widely, and hold large codebases, is still unclear. Supporters have long described formal verification as the closest thing to an ideal in software security, because it uses mathematics to check that software behaves as intended — and therefore, they argue, whether it has critical bugs.
The significance is methodological: it turns “no bugs” from something humans review repeatedly into a proposition a machine can check. The hard part is not getting the first proof through, but re-proving it after every change, or the proof expires along with the code. So the question to watch is whether this becomes a routine step before commits rather than a one-off demonstration.
DeepMind ships Gemini Robotics 2.5 and publishes the ASIMOV-jailbreak safety benchmark
Over to robotics, where the point this time is not a demo video but actual factory work.
Google DeepMind released a new generation of its AI robotics platform, Gemini Robotics. Google says that at 2.5 the system’s ability to complete real-world tasks is close to human level. Some Gemini robots have already done real work inside Mercedes-Benz factories. The team also published an open safety benchmark, ASIMOV-jailbreak, as preparation for robots and people working side by side. The stack is layered between a user’s instruction and the robot’s motion: a vision-language model, Gemini Embodied Reasoning, plans the task from the instruction and live video; an actuator-generating vision-language-action model — Gemini Robotics, now at 2.5 — produces the motion. DeepMind says 2.5 is considerably better at tasks that need high dexterity. Evaluations cover more than 200 activities, including dispensing food with tongs and using a pipette the way a lab does. Google also showed Apptronik’s humanoid Apollo 2; with Gemini Robotics 2.5 it can tie a bow with a five-fingered hand — deformable objects like ribbon have long been hard for robots.
The real change is the division of labour: planning goes to a language model, motion to a dedicated action model, and the human supplies a single sentence. That split lets capability improve separately, and makes it easier to locate which layer failed. What to watch is whether the public benchmark holds up against targeted attacks, and whether the same level survives in environments nobody has pre-arranged.
Meta AI releases Muse Ember 1.0 for the short tasks agents repeat
Next, an engineering question: where exactly a small model gets cheap.
On November 10, 2026, Meta AI launched Muse Ember 1.0. Meta calls it its smallest and most efficient model, aimed at the tool calls, coding subtasks, routing, extraction and long-document parsing that agents repeat. Independent verification in the source material is still thin. The model takes up to one million input tokens. Below 200,000 prompt tokens, input, cached input and output cost 0.10, 0.004 and 0.50 dollars per million tokens. Above 200,000 tokens, those rise to 0.30, 0.012 and 1.50.
Within their short-prompt bands, Ember matches GPT-6 Luna and Claude Haiku 5.5 at 0.10 dollars input and 0.50 output per million tokens. Claude Fable 5.2 is 8 and 40 dollars. The long-prompt surcharge thresholds are 200,000 tokens for Ember, 272,000 for GPT-6 Luna and 100,000 for Claude Haiku 5.5; the latter two charge 0.20 and 0.75, and 0.50 and 2.50 dollars, on long prompts. Ember offers six reasoning tiers from minimal to max. A chart from Artificial Analysis that Simon Willison cites puts xhigh and max at roughly 0.12 and 0.14 dollars per task, an Intelligence Index around 34 and 36, and on the Pareto frontier. That is one specific benchmark. In the same observation the high tier did not return a complete SVG drawing, and the reasoning tiers did not line up cleanly with visual quality.
This release gives a million-token ceiling, a 200,000-token price threshold and six reasoning tiers at once. Agent builders can now cost short prompts and long-document work separately. The evidence covers official positioning, one benchmark chart and one SVG observation — not enough to claim it is cheaper on real workloads.
The practical use: bill short prompts and long documents separately, and the cost structure becomes legible, which makes it easier to decide which tier gets which task. But the evaluation evidence is thin — official positioning, one chart, and one incomplete drawing. So do not put it in the budget yet; run your own call patterns for a week and see whether the savings outweigh the quality loss.
Glasswing reported 37,966 vulnerabilities in seven months; only 217 are marked fixed
Then a less flattering number, about the distance between finding and fixing.
Anthropic’s Glasswing experiment ran for seven months and reported 37,966 vulnerability findings. Only 7,349 became public — 19.4% — and only 217 are marked as fixed. The gap between AI agents finding vulnerabilities and driving fixes is stark. A flood of reports can also add to the burden on open-source maintainers to verify, deduplicate, withdraw and patch. Only part of the total is public, and that part includes withdrawals and duplicates; the rest is embargoed or not yet disclosed, with truth and validity unclear. The data therefore cannot establish whether AI’s net effect on software security is positive or negative.
Project Glasswing is a multi-party vulnerability research programme meant to use Anthropic’s frontier model Mythos to find and fix software flaws before attackers exploit them. A public disclosure ledger records public, duplicate, withdrawn and fixed states. So the published CVE count is only part of all findings, and cannot be read as the number confirmed or fixed. For open-source maintainers and the organisations depending on those projects, the gap between discovery and repair becomes work: verifying, deduplicating, writing patches and shipping them. As of May 22, 2026, Anthropic had disclosed 1,596 vulnerabilities to 281 open-source projects. An undisclosed finding is not the same as an unfixed or invalid one.
Commenters accept that the ratio between findings and fixes deserves attention but question how much it explains: public samples contain withdrawals and duplicates, undisclosed parts may still be embargoed, and public results cannot be extrapolated to all findings. Others worry the fix side is underfunded, or that patches fail to cover older versions, citing PHP2shell’s maintenance to show how hard landing patches is. Others note that the article’s conclusion sits close to VulnCheck’s commercial interests and stay wary of that position.
What this pinpoints is the most fragile link in automated disclosure: once finding bugs is nearly free, the pressure lands entirely on maintainers. The disagreement in the comments is worth hearing too. What to watch is not how many findings there are, but whether disclosure pace and repair investment keep up — otherwise the reports become an inventory nobody works through.
Blender 5.3: Gaussian Splatting support
Then a tooling update, this time for people building 3D scenes.
Blender 5.3 is out with support for Gaussian Splatting, plus extended Geometry Nodes bundle capabilities, so working with splat data and fairly complex visual programming is more direct. One early tester tried a Spirula Studio export containing four million splats on an M4. They report it runs fast, that the integration is mature, and that they found no inconsistency with the rest of Blender’s interface. That suggests real potential for large single scenes, but the feedback is not enough to show broad industry impact.
Gaussian Splatting represents a scene as a set of 3D Gaussian ellipsoids with position, scale, orientation, colour and opacity, composited by projection into an image, without converting to a conventional triangle mesh first. Official development notes show 3D Gaussian Splats wired into import, point clouds and Geometry Nodes, and into the Workbench, EEVEE and Cycles renderers. Geometry Nodes is Blender’s node-based procedural geometry system; bundles group related values into a unit that works with structured data like lists and closures. With native support, users can import, inspect and render high-detail scenes reconstructed from photos or video inside Blender, joining the existing pipeline.
Early users broadly accept the speed and maturity. Some developers argue bundles, lists and closures meaningfully strengthen complex workflows, at the cost of learning the programming model. Others propose triangle-based splatting, arguing GPUs already handle triangles well and that it may suit real-time games better.
For 3D practitioners, the value is skipping an export and conversion detour and putting reconstructed scenes straight into the existing pipeline. But one early trial cannot carry an industry-level conclusion, and heavy scenes still need hand optimisation. What to watch is whether the plugin ecosystem migrates, and whether triangles or this representation win in stricter settings like real-time rendering.
Taken together today: model capability keeps climbing, while pricing, trust and the cost of repair are all on the table at once. How those accounts balance out will get answers over the next few weeks.
🎧 This episode is also available as a podcast: listen to Daily AI Stew · 2026-11-11.