TODAY IN 30 SECONDS
● Anthropic Cyber Mission — 11 founding partners, free Claude scans for open-source projects.
● Nano Banana 2.1 — image prices halved, predecessor dies 29 October.
● Manus raises $500M+ after Beijing blocked Meta's reported $2B buyout.
● CVE-2026-105192 — unpatched CVSS 9.8 in LMCache. If you run vLLM, read item four.
1. Anthropic launched a cybersecurity programme
The Cyber Mission went live on 8 October with 11 founding partners — Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation. The open-source half, OSS Scanner, gives enrolled projects free periodic scans from Anthropic's most capable models, each report carrying a proof of concept and a suggested fix.
The caveat is in the announcement: reports are model-generated and sent without human review, and the expected true-positive rate is above 90% — meaning roughly one in ten is wrong.
Full story and how to enrol →
2. Google halved its image prices and set a deadline
Nano Banana 2.1 went generally available on 6 October at exactly half the per-image cost of Nano Banana 2 at every resolution: 1K from $0.067 to $0.0336, 2K from $0.101 to $0.0504, 4K from $0.151 to $0.0756.
Two things most coverage buried. Nano Banana 2 shuts down on 29 October, 23 days after launch. And the 512px tier is gone entirely — 2.1 does 1K, 2K and 4K only.
Full pricing table and migration notes →
3. Manus parent raised more than $500 million
Butterfly Effect closed the round on 8 October, led by Boyu Capital and IDG Capital with Tencent, HSG and ZhenFund participating. It follows Chinese regulators blocking Meta's reported $2 billion acquisition. A roughly $4 billion valuation has been reported but is unconfirmed.
Full story →
4. An unpatched CVSS 9.8 in LMCache
JFrog disclosed CVE-2026-105192, an unauthenticated remote-code-execution flaw in LMCache, the KV-cache layer commonly deployed alongside vLLM servers. Affected versions run from 0.3.9 through 0.5.5, which is the current stable release — and at the time of disclosure no fixed version existed.
If you self-host inference on vLLM with LMCache, this is today's actual priority. Until a patch ships, the standard mitigation applies: make sure the LMCache service is not reachable from untrusted networks and sits behind authentication at the network layer. Check your deployment rather than assuming a default is safe.
5. Also worth knowing
- StepFun released Step 5 Preview — a 600B sparse MoE with 27B active parameters per token and a 1M context, at $1.00 input and $2.70 output per million. Open weights are promised for 15 October.
- Anthropic updated its usage policy, adding prohibitions on sustained abusive behaviour toward Claude and new sections on democratic processes and deceptive campaigns such as fake accounts and fabricated news outlets.
- Atlassian and OpenAI expanded their partnership on 6 October, connecting Rovo and the Teamwork Graph to OpenAI models and putting Jira, Confluence and Bitbucket context inside ChatGPT and Codex. Atlassian says more than 3,000 of its own developers use Codex.
- Arena launched an Alignment Index scoring 27 models on agent failure signals, with GPT-6.1 Sol leading at 87.9, and disclosed a $200M Series B at roughly a $3.1 billion valuation.
- Liquid AI released d1-3B, an open-weights decision model that returns typed answers in a single forward pass with zero output tokens.
The thread running through today
Three of these are price or capital stories pointing the same direction. Anthropic cut Haiku's input price 90% on the 7th, Google halved image generation on the 6th, and an agent company raised half a billion dollars into that falling cost curve on the 8th.
The models are getting cheaper faster than the products built on them are getting better. That is a good environment to be building in and a difficult one to be raising against, and this week showed both.
Sources