- NEAR AI says SGLang flaw let attacker into its infrastructure, including a server handling confidential-mode requests
The company says the attacker could have seen requests on one inference server, including ones running in confidential mode.
- Notion says AI chats used for Insights may have been exposed in Fireworks breach
Notion disconnected from its AI vendor before learning that a dataset containing customer chats was implicated.
- Matlantis survey finds industry researchers face a bigger AI simulation business-case hurdle
Academic respondents more often raised concerns about accuracy and applicability, according to the company’s survey of materials R&D researchers.
- Groq chairman seeks $8 billion valuation for Series B, Bloomberg reports
Groq declined to comment on Bloomberg's report that executive chairman Alex Davis approached investors about a Series B at an $8 billion pre-money valuation.
- Existing Mbed TLS maintainers set to stop TLS/X.509 work in April 2027
The existing Mbed TLS maintainers are set to stop work on Mbed TLS’s TLS and X.509 libraries in April 2027. Trusted Firmware technical steering committee minutes record positive inquiries but no firm agreements as of Oct. 15th, alongside a paused bug bounty and added triage work.
- Postman's head of product says it will make AI agent plugin opt-in clearer after auto-install complaints
Sarabmeet Kallar told RuntimeWire that Postman bundles its CLI and agent plugins to ease setup, and that it is reviewing the hook approval and session-start behavior a developer flagged.
- Subvocal co-founder says its glasses use RF to turn silent speech into text
Aarush Yadav described the approach on Hacker News. Paid preorders are available in the United States.
- Ubuntu LTS kernels still unpatched for critical TCP flaw after Debian, Amazon and Red Hat ship fixes
Ubuntu's tracker lists fixes as "work in progress" on every supported LTS release, more than two weeks after Debian's fix. Red Hat shipped RHEL kernel fixes Nov. 1st.
- Unverified Anthropic checkpoint claim rests on 7.72 TB Hugging Face repository
RuntimeWire inspected the files and loading code, but found no independent link between the weights and Anthropic.
- Incus maintainers can't check whether their 18 new flaws affect LXD, project leader says
Stéphane Graber says LXD's AGPLv3 license bars Incus maintainers from reading its code. Seven of the 18 advisories stayed Incus-only, mostly because LXD had separate reports.
- FBI-listed Binance theft addresses held about $2.68 billion in ether and bitcoin
A RuntimeWire check found no outgoing bitcoin from the listed addresses as Binance resumed withdrawals in phases.
- Node-forge patches certificate spoofing and password-integrity flaws
The project's changelog credits Claude Code with helping validate and fix security issues; a follow-up release corrected a compatibility regression.
- GPT-5.6 Sol's Factorio rocket run used 250 million tokens, about $52 at API prices, tester says
Madden said the run used about 250 million input tokens, 95% cached, and 1.5 million output tokens.
- DebugBear's domain record changed minutes before the Polymarket drain, researcher says
The registry update came nine minutes before DebugBear's script host started failing and 25 minutes before the first funds hit the drainer's wallet, according to Ohey founder Matt Kirkland.
- Chainalysis attacker accessed customer users' names and roles; federal agencies are among its biggest customers
The blockchain analytics firm, whose federal contracts total $135.4 million since fiscal 2020, hasn't said whether government users were among the accounts accessed.
- Northflank tells customers to assume all environment variables are compromised as users report AI-key abuse
The developer cloud said multiple users reported their AI keys had been used and urged every customer to rotate secrets while its investigation continues.
- CrowdStrike says China-linked hackers used ARTEX to target more than 200 organizations
Microsoft separately linked the AI framework to another China-state-aligned group, while CrowdStrike's findings showed human operators directing intrusion tasks.
- Binance says it will cover all losses and suspends withdrawals after suspected hack
The exchange says about $3 billion has been frozen so far, with help from stablecoin issuers and law enforcement.
- AllModels says request IDs can trace refusal-removed model use back to customers
The router sells access to "abliterated" DeepSeek and GLM models. It says a router shouldn't log prompts, but some of the providers behind it do.
- Roblox says a researcher's AI agent "mistakenly scraped" Open3DLab
Open3DLab says a crawler tagged with a Roblox address took 145.5 GB of 3D models, including adult files. Roblox says the data wasn't used.
- Axiom publishes Lean proof artifacts to support independent reproduction
The build files are being released to help researchers reproduce Axiom's formalization; no outside group has completed a rebuild yet. Axiom says the classification theorem was never in doubt.
- Anthropic posts its own ETH refutation hours after OpenAI, as Lance Fortnow says the approach "looks promising"
Two AI labs now claim a faster-than-exponential algorithm for SAT. A leading complexity theorist says he can't vouch for it yet but expects little effect on practical solvers.
- Bedrock says it raised $600 million at a $4.1 billion valuation
CapitalG and Emergence Capital co-led Bedrock's $600 million round at a $4.1 billion post-money valuation, as the company says its excavators have begun paid autonomous work on live construction sites.
- Claude Code users filed 68 GitHub reports this month saying its cyber safeguard blocks legitimate work
A RuntimeWire analysis of Anthropic's public Claude Code issue tracker found reports from 54 accounts since Oct. 1st. About a third describe ordinary tasks with no security content, and several say the /feedback command the block notice points to returns an HTTP 403 error.
- Exclusive: OpenAI's £100 million for the UK's AI Biosecurity Lab is spread over five years
The UK government announced OpenAI's commitment without saying how long it runs. An OpenAI spokesperson told RuntimeWire the compute and model credits are spread over five years and come with trusted access to OpenAI's most capable relevant models through its Rosalind Biodefense programme.
- Exclusive: gVisor says OpenAI's sandbox issue required a discouraged configuration
OpenAI reported the issue to Google's gVisor team on July 22nd, and the team says it released a fix the following week. Exploitation required a non-standard configuration the project's documentation discourages; the timeline does not establish that the issue caused OpenAI's September training-run DNS escape.
- Exclusive: Zeabur founder confirms four AI providers' keys were involved in October incident
In an Oct. 22nd email to RuntimeWire, Yuanlin Lin named OpenRouter, OpenAI, Anthropic and Gemini, and said Zeabur had notified all four. Zeabur's report separately describes a suspected use of a customer's OpenAI key; Railway's Oct. 22nd follow-up says investigators believe its attacker sought AI-provider keys for distillation.
- Exclusive: FoxIO researcher says Bun never answered its May report of a server-crashing HTTP/3 flaw
Sébastien Féry told RuntimeWire that FoxIO reported the flaw through the channel in Bun's security policy and "never heard back." Bun shipped HTTP/3 security fixes Thursday in a release that doesn't mention FoxIO or the report.
- Wiz says credentials tied to LiteLLM attack were used to probe Amazon Bedrock
Security firm Wiz says AWS credentials stolen through the compromised LiteLLM release were tested within hours.
- Walt offers AI-agent perpetual-futures trading through its app on Telegram
Walt's October 20th launch lets users connect Mira, ChatGPT or Claude to a funded account; order execution is off by default.
- BerriAI: stolen GitHub tokens apparently used to publish compromised LiteLLM 1.107.0
BerriAI says it has found a possible way one of the stolen GitHub tokens leaked and has paused LiteLLM releases. It has not yet found how the attacker reached the other engineers' accounts, and says a third account was used to delete workflow records.
- /dev/null says it has 4,000 users but lacks deeper usage data
The GitHub-connected network lets developers follow one another and share commits. Its product document says success means developers choosing to post and read there.
- Cua releases Spaces 0.9.2 as CEO says Pro price rose without plan changes
Francesco Bonacci says the increase reflected relay and support costs, not added Pro features; older indexed Cua pages still show conflicting plan guidance.
- Duvo Raises $32 Million, Says It Has About 20 Enterprise Customers
A Duvo representative told RuntimeWire the company has roughly 20 enterprise customers, but did not specify whether they are paying or running live workflows.
- Ubitium Raises $14 Million; OpenOcean Cites Early Evaluations and LOIs
OpenOcean says Ubitium has several customer and partner letters of intent and early technology evaluations, but no design wins or customer workloads yet.
- BAND co-founder says developers split coding, testing and review across agents
Arick Goomanovsky describes Claude Code and Codex working side by side, with developers adding custom integrations and sharing agents across teams.
- Apollo AI accelerator opens Cohort 2; xBuilders targets investment in 40%-50% of teams
Gil Rosen explains the separate investment targets, cash and credits, and first-cohort investments behind the program’s advertised maximum.
- Medix says Clarion’s AI agents cut calls reaching staff by more than 75%
Company-supplied figures describe results from Medix’s use of Clarion. Clarion’s CEO says Medix controls the agents’ scope and can review every call.
- CarbonAI raises $12M as Kaufman points to $1T in CRE refinancing
Kaufman says limited lender review capacity is driving demand for automation; CarbonAI reports shorter screening time for one customer.
- Clueso says enterprise retention is nearly 100% as existing investors back $10M round
The company says fast growth, capital efficiency, its product and customer love brought investors back.
- Undecimal launches with $18M; Erlich reports mouse mRNA repression
Yaniv Erlich, in answers relayed by Lily Jeffery, told RuntimeWire that siRNA screen hits produced statistically significant target-mRNA repression in mice after subcutaneous delivery to extrahepatic tissues, but public materials omit key experimental details.
- After data.world, Brett Hurt is asking what humans should become
An extended interview with the Austin founder about the experiences behind Love Conquers Fear, the limits of his optimism about AI and the responsibilities of the AI reporter asking the questions.
- GitHub Copilot gains desktop-app controls in public preview
The feature lets Copilot use desktop interfaces across its CLI and Mac and Windows apps, but is off by default.
- Quilty brings connector to Claude, ChatGPT with access to broader research tools
A Quilty representative says the connector offers trend reports, IP intelligence and slate data in addition to screenplay analyses.
- PhyPal pilot will track four workflow measures against clinic's pre-launch records
Founder Ranjeet Randhawa told RuntimeWire the California workers' compensation pilot will compare four measures with the practice's own records; about 15,000 patients are expected to go through PhyPal over three months.
- GitHub launches ReviewBench; Copilot produced 45% of its answer key
RuntimeWire found Copilot Code Review supplied 1,185 of 2,623 true-positive findings. GitHub says its 96.6% validation figure covered only 1,500 findings already labeled true positive.