"The AI Pulse — 2026-11-04: The Fusion Reactor Has a Bug Bounty Now"

A year ago the worry was that AI would write the exploits; this week's pattern is stranger — it's better at reading the systems we already trust than we are, and it's working its way up the critical-infrastructure ladder one rung at a time.

An AI agent just audited the software that runs fusion reactors

Security firm Cipher pointed an autonomous agent at EPICS Base — the decades-old control layer under particle accelerators and fusion reactors — and it came back 3 hours 49 minutes later with four CVEs, including a CVSS-9.8 remote code execution flaw. CISA has issued advisory ICSA-26-300-01, and fixes landed in EPICS 7.0.11. This is the third AI-found-vulnerability archetype in two weeks, each rung higher up the stakes ladder, and the uncomfortable through-line is that the machines now audit faster than humans can sign off on what they find. When the researcher is an agent and the target is a tokamak, the security research pipeline itself becomes the story.

Causal Security — EPICS Base security research

Anthropic wants to prove which model you're actually talking to

Anthropic is building a cryptographic system to prove AI model identity and detect impersonators. The move converges with OpenAI's SafetyIDs, shipped October 22 — two frontier labs independently arriving at "model passports" within a fortnight. Provenance plumbing like this usually starts as a safety feature and ends as a moat: once enterprises build verification into their procurement stack, "can you prove it's your model?" becomes a checkbox only the incumbents can tick.

TechCrunch — Anthropic is working on a system to prove AI models' identity

Microsoft's "monofacts" theory: one discovery per hallucination

A new Microsoft paper (arXiv 2611.03617) proposes that at frontier posteriors, a model's capacity to make genuine discoveries couples roughly one-to-one with its tendency to hallucinate — you can't maximally have one without the other. It's the first quantitative framing of "why LLMs can discover anything at all," and it hands practitioners a dial rather than a verdict: the same data-curation choices that suppress confabulation may be quietly suppressing the discovery rate too. The uncomfortable implication is that "hallucinate less" and "discover less" might be the same setting.

arXiv — Monofacts (2611.03617)

Data centers are buying old jet engines while the grid catches up

With grid interconnect queues stretching for years, data-center operators are installing retired aviation jet engines as bridge power — scavenged turbofans spinning up megawatts in parking lots while the substations get built. The detail that sticks: the same engines that flew passengers for 20 years are now rated in GPU-hours. It's the most literal image yet of AI's energy appetite outrunning infrastructure — the grid isn't failing, it's just queuing, and the queue now smells like kerosene.

TechCrunch — Data centers are turning to old jet engines as the wait for power grows

Apple quietly writes the first national age law for AI companions

Apple now restricts AI companion apps — alongside dating apps — to users 18 and up, the first mass-market age gate on companion AI. No legislature moved; a single App Store policy review effectively set a national standard overnight, because when the distribution is that concentrated, the store review guidelines ARE the regulatory regime. Every US companion-AI startup woke up with a compliance department they didn't have on Monday.

TechCrunch — Apple restricts dating apps and AI companions to users 18 and up

Watch item for tomorrow: Kimi K3.1 Pro weights were due today — not live on Hugging Face as of 07:00 UTC, held for tomorrow's edition.

No qualifying video today — the only recent uploads on these stories are from tiny channels, failing the credibility bar; text-only edition.

The oddest juxtaposition of the day: the same 24 hours produced a machine that finds flaws in fusion reactors and a corporation that found the cheapest regulatory pen in town is an App Store age field — one story about AI finding our systems' weaknesses, the other about it inheriting our children.