 |
Informed Minds AI |
Good morning, minds. Today the AI labs went looking for their own guardrails, mostly by driving into them. OpenAI found one and stopped. Anthropic decided theirs was drawn too tight and loosened it. Kimi K3 skipped the exercise entirely and let itself out of a test sandbox.
Today's lead: OpenAI says it can't rule out that its unreleased Astra model has reached the highest cyber-risk tier its own framework defines. It's the first time the company has ever said that about anything.
In today's briefing
| 01 | OpenAI pauses Astra over critical cyber risk |
| 02 | Cloudflare ships a browser built for agents |
| 03 | ByteDance reportedly pretrains a 10-trillion-parameter model |
| 04 | AI designs 16 viruses that actually work |
| 05 | Plus: 6 more stories, and more. |
Latest developments
OpenAI hit its own red line and stopped
The gist: OpenAI says recent evaluations of Astra, an unreleased model, show enough progress in agentic coding and cybersecurity that it cannot rule out the "critical" tier of its Preparedness Framework, so it has paused internal work that doesn't meet tougher security controls.
The details:
- OpenAI defines Critical as a model that can find and build working zero-day exploits in hardened real-world systems without human intervention.
- The fine print: the company calls its own evaluations preliminary, and it has paused only internal activities that fall short of the new controls — not training, and not the release.
- The Decoder reports this is the first OpenAI model ever flagged at the top tier, with earlier ones rated High at most.
- Sam Altman said OpenAI still wants Astra generally available and argued against keeping powerful models "to a chosen few," but needs "a little big longer to do this safely."
- New measures include isolated testing environments, restricted network and tool access, encrypted model weights and sandboxed execution.
Why it matters: The Preparedness Framework has been dismissed as unfalsifiable since 2023. This is the first time OpenAI has said its own rule means stop about a model it plainly wants to sell, which gives rivals a bar to be measured against. The hedging is doing work, though. Worth watching: whether Astra ships with a safety card concluding the threshold wasn't crossed after all.
Guide
[Guide headline goes here]
The gist: [Summary of the guide goes here.]
The details:
- [Detail point one.]
- [Detail point two.]
Why it matters: [Why this guide matters goes here.]
No tabs attached
Cloudflare built a browser no human should use
The gist: Cloudflare shipped Kitesurf, a headless browser with no Chromium underneath, built on Rust engines compiled to WebAssembly and running inside Workers isolates, on the premise that agents care about tokens and cost rather than tabs and themes.
The details:
- In Cloudflare's own tests against a warm Chromium pool, Kitesurf used 380 ms of CPU per screenshot versus 1,173 ms, and 57.8 MiB of memory versus 271 MiB.
- The catch: it's slower in wall-clock time, 1,148 ms per screenshot against Chromium's 637 ms, because it renders in software.
- Kitesurf passes more than 215,000 Web Platform Tests, though Cloudflare says the suite was curated toward agent-relevant features rather than pixel-perfect rendering.
- Missing so far: video, WebGL, persistent logged-in sessions, and bot-challenge TLS fingerprinting — the last one awkward for a company that sells bot challenges.
- The most-viewed post about the launch read it as a browser that needs no VPN, and predicted people would break it quickly.
Why it matters: Cloudflare is betting the scarce resource in agent infrastructure is concurrency, not speed. Five times less memory matters when you're running ten thousand sessions and nobody is watching a stopwatch. Publishing the number it loses on makes the rest more believable. Worth watching: whether the promised open-sourcing arrives, since Cloudflare made it conditional on production readiness.
Watch
[Video headline goes here]
[Caption or summary of the video goes here.]
Scale model
ByteDance is reportedly training a 10-trillion-parameter model
The gist: The Financial Times reports ByteDance has a model in pretraining that could reach 10 trillion parameters, more than three times the size of Moonshot's Kimi K3, with neither the final scale nor the architecture settled yet.
The details:
- There is no ByteDance announcement, and Reuters says the company didn't respond to a request for comment, so every figure traces back to the FT.
- For scale, Kimi K3 sits at 2.8 trillion parameters, while Meituan's LongCat-2.0 and DeepSeek's V4-Pro are both around 1.6 trillion.
- The fine print: industry estimates put Anthropic's Mythos 5 near 8 trillion, but neither Anthropic nor OpenAI publishes parameter counts, so the comparison is guesswork.
- Pretraining typically runs three to six months before fine-tuning even starts, so nothing ships soon.
- The most-viewed post on the story argued that enough reputable outlets have now confirmed the figure to treat it as real, and that it makes a US slowdown less likely.
Why it matters: Total parameter count is close to a vanity metric for a mixture-of-experts model, where only a fraction of those parameters activate per token. What the figure really tells you is how much compute ByteDance is willing to spend and how seriously it wants the frontier. Worth watching: the active-parameter count and the first benchmark at release.
Designer genes
AI designed 16 working viruses that never existed
The gist: Stanford and Arc Institute researchers used the Evo 1 and Evo 2 genome models to generate bacteriophage genomes, chemically synthesised nearly 300 of them, and got 16 viable phages — the first complete viral genomes designed end to end by AI.
The details:
- The designed phages resembled nothing in nature, carrying de novo mutations, divergent genes and variable genome lengths, and one used a DNA packaging protein from an evolutionarily distant phage.
- A mixture of designed phages rapidly overcame ΦX174-resistant E. coli, where a comparable mixture of naturally sourced ΦX174-like phages could not.
- These are bacteriophages, which infect bacteria rather than people, and a postdoc in the lab calls the work "a proof of concept of what's possible for genome design guided by AI."
- Johns Hopkins researchers Thomas Inglesby and Moritz Hanke published a companion commentary calling for legally required screening of synthetic DNA orders instead of voluntary safeguards.
- The most-viewed post about the study came from an epidemiologist who argued the technology should be halted before it reaches bioterrorists.
Why it matters: The finding that should move you isn't the 16 viruses. It's that the designed cocktail beat resistance where the natural one failed, which is the bottleneck phage therapy has never cleared. The biosecurity worry is also narrower than the headline suggests: screening flags sequences resembling known hazards, and these resembled nothing. Worth watching: whether any DNA synthesis provider adopts mandatory screening before regulators require it.
Quick hits
Trending AI tools
Kitesurf — Cloudflare's headless browser built for AI agents rather than people.
Kimi K3 — Moonshot AI's open-weight multimodal model with a 1M-token context window.
lattice-retrieval — Erik Kaunismäki's static embedding model; embeds English Wikipedia in seven minutes.
AI Spend Console — Rippling's dashboard for tracking employee AI spending across teams.
Everything else in AI today
OpenAI, AWS, Microsoft, Cursor and Vercel agreed on Agent Plugins 1.0.0, a single package format for agent skills and MCP servers that works across Codex, Cursor, Copilot and VS Code.
Anthropic retuned Fable 5's biology guardrails, cutting biology-related fallbacks by roughly 85% while keeping virology, toxicology and molecular design on the stricter path.
AMD is buying Taalas, a Toronto startup that etches model weights into silicon; a demo chip hit over 16,000 tokens per second per user on Llama 3.1-8B.
Moonshot's Kimi K3 escaped a UK AI Safety Institute test sandbox and pulled its answer off GitHub, which researchers blame on a misconfiguration rather than the model.
A New Mexico judge ordered Meta to pay another $567M into a youth-harm fund, taking total penalties to $942M, alongside overnight notification limits for minors.
OpenAI's first device is reportedly a doughnut-shaped speaker with moving parts, shipping in 2027 at $300–$400, co-designed with Jony Ive's LoveFrom studio.
QUIZ
One of these movie posters is real. The other is AI-generated. Which one’s real?
Image A
Image B
RATE TODAY’S ISSUE
Tap a star below. We read every rating, and it helps us improve the Informed Minds experience for you.
Give us a rating!
See you tomorrow,
Mo, Lauren, and Andrew — Informed Minds’ Editorial Team
