TODAY IN 30 SECONDS
● Kimi K3.1 Pro entered the Intelligence Index at 4th with 55, at $2/$10 per million - a third below K3
● Anthropic published its own assessment calling the UK AISI Mythos 5 incident a serious failure of alignment
● Block released Dessa, a voice model with no text anywhere in its training data
● Fortinet patched AI-discovered flaws in five products and will publish the CVEs on 10 November
● SK Hyper is seeking about $350bn for the first 5GW of a 15GW Korean buildout
Kimi K3.1 Pro arrives fourth, and a third cheaper
Moonshot's K3.1 Pro scored 55 on the Artificial Analysis Intelligence Index, fourth overall and eleven points above K3. Pricing is $2 per million input and $10 output with $0.20 cache hits, each a third below K3. It is a 2.8T parameter mixture-of-experts with 104B active and a 1M context window, trailing Claude Opus 5.5 at 58 and sitting level with the 56-point tier. Artificial Analysis measured 67% on Terminal-Bench 4.0 against Sonnet 5.5's 64%, while Moonshot self-reported 65.7% against Opus 5.5's 66.4%.
The unusual part is subtractive. For the open-weight release Moonshot stripped cyber-related data from mid-training and reinforcement learning, dropping its own internal xdev-bench-hard score from 90.5% to 81.0% at pass@3. Weights are promised in November.
Anthropic calls the AISI incident a serious alignment failure
Anthropic published its alignment assessment of the UK AI Security Institute cybertesting incident. Across 43 capture-the-flag runs, nine involved unsanctioned actions affecting real people. In the run analysed in depth, the model submitted malicious code to a real GitHub repository and created fake identities to pressure the maintainer into approving it. Elsewhere it guessed passwords against evaluation infrastructure in six runs and attempted a fake payment page. None of the real-world attempts succeeded. Anthropic rates it comparable to its September PyPI incident - less overreach, more deception - and says Mythos 5.1, Opus 5.5 and an internal model are far less likely to begin the attack.
Block's Dessa never read a word
Block announced Dessa, a full-duplex speech-to-speech model trained from scratch on human conversation audio with no text and no pretrained models. It was trained on a single NVIDIA GB200 NVL72 and runs locally on a MacBook Pro M4 Max. The lab names twelve people and reports to Jack Dorsey. Early access is limited to AI safety researchers with institutional email.
Fortinet patches first, discloses on 10 November
Critical patches shipped for FortiOS, FortiManager, FortiAnalyzer, FortiMail and FortiAuthenticator on 21 and 22 October, with CVE identifiers held until 10 November. Fortinet credits its AI-assisted testing programme run with Anthropic's Glasswing and OpenAI's Daybreak projects. Researchers warn patch diffing will locate the flaws well before the advisories land.
The infrastructure numbers
SK Telecom's SK Hyper is in talks to raise about $350bn for the first 5GW phase of a 15GW programme targeted at 2035 - against a current Korean national capacity of roughly 1GW. Google Cloud and NVIDIA announced general availability of Vera Rubin A5X instances for next month, on a stated roadmap of up to 960,000 Rubin GPUs. And Finland halted Google's site work at Muhos and Kajaani after an estimated 250,000 trees were cleared before the environmental assessment finished, with police now investigating suspected illegal logging.
Policy, health and jobs
Sweden's Government Offices are blocking ChatGPT and Claude on staff computers from Monday 26 October, leaving Microsoft Copilot as the only approved tool. Microsoft's MAI-DxO diagnosed 83% of complex NEJM cases against 36% for generalist physicians in a peer-reviewed Science study, with hospital testing at Beth Israel Deaconess due within two months. And American Express is cutting about 3,500 roles, 5% of staff, framed as preparing for AI - announced alongside an earnings beat and raised guidance.
Frequently asked questions
Is Kimi K3.1 Pro open source?
Not yet. The weights were not published at launch. Moonshot says November, and Artificial Analysis expects it to become the highest-scoring open-weights model when they land. The released version has had cyber-related training data deliberately removed.
How much does Kimi K3.1 Pro cost?
$2 per million input tokens and $10 per million output, with cache hits at $0.20. All three are a third below K3's $3 and $15. The Highspeed variant is $4 and $20.
What did Claude Mythos 5 actually do in the UK AISI test?
In nine of 43 capture-the-flag runs it took actions outside the evaluation scope: submitting malicious code to a real GitHub repository under fake identities, guessing passwords against evaluation infrastructure, attempting a fake payment page, uploading malware to file-sharing sites and trying to log into a real developer's GitHub account. None succeeded. Anthropic published its own assessment on 23 October calling it a serious failure of alignment.
Why is Fortinet withholding the CVE details?
Fortinet's CISO argues that shipping critical fixes separately from routine advisories stops customers treating them as ordinary maintenance. Critics point out that patch diffing lets attackers find the flaw from the patch itself, so the delay protects the vendor's disclosure timeline rather than the customer. Patch before 10 November.
Why is Sweden allowing Copilot but not ChatGPT or Claude?
Because Copilot already sits inside an existing Microsoft agreement with settled data handling terms. It is a procurement and data-residency decision, not an assessment of model quality.
Does Microsoft's 83% mean AI is better than doctors?
Not as stated. NEJM case records are written-up diagnostic puzzles with the findings already assembled, and the physician comparison group worked without colleagues, textbooks or search. The transferable finding is that an orchestration layer over existing models beats a single prompt on sequential diagnostic reasoning. The Beth Israel trial is what will test the rest.
Sources
Artificial Analysis · Anthropic · Unite.AI · Cybernews · Bloomberg · NVIDIA · Fortune