Browsers blocked eight weak CA keys within hours. Chrome drops the root on Oct 27, and eMudhra is quitting the global TLS business.
Eight intermediate certificate authority keys at eMudhra, an Indian digital-trust company, were numbers below 256, Mozilla’s Dennis Jackson writes in Bugzilla, so anyone could work them out. Whoever holds such a key can sign a certificate for any website. A researcher named Wayne reported it to Apple, Google, Microsoft, Mozilla and eMudhra at 18:05 UTC on Saturday, October 17. Google says Chrome blocked the keys at 19:45 UTC, and Mozilla finished its Firefox block at 22:59 UTC.
From Mozilla Bugzilla 2083449, Google Security Blog, Mozilla Security Blog, eMudhra status page, eMudhra exit announcement, Mint (CERT-In advisory), The Hindu BusinessLine, University of Michigan Safe ComputingRead the story →
What fits in 256 GB and 384 GB, how fast it runs, how close the quants stay to full precision, and which inference engine to pick for each build.
Two Sparks run GLM-5.3-Flash at 4 bits or DeepSeek-V4.1-Flash-1017 under 3 bits, at 60 to 65 tok/s on the best builds. Three run DeepSeek’s official weights or the full GLM-5.3. Measured speeds, quality checks, engines, other hardware and what we could not verify.
Sourced from builders’ posts, recipe repos, model cards and NVIDIA documentationRead the feature →
For 18 months, eight certificate authority keys could be guessed in seconds. Browsers blocked them within hours. Now 30,800 certificates have to move before Chrome cuts the root on Oct 27.
What a certificate authority is, how browsers pull trust, how long the weak eMudhra keys were exposed and who has to act. Includes a timeline, a map of what is and is not affected, a five-minute check and a list of what is still unconfirmed.
Sourced from our feeds, vendor advisories, government alerts and news reportsRead the feature →
With @theoremlabs, over a few weeks. Formal verification used to need small codebases or a lot of expert time.
Anthropic’s Frontier Red Team says it used Claude to formally verify software at large scale. Formal verification can prove code is free of certain bugs, Anthropic says, but it has only been feasible for small codebases or with a lot of expert time.
Three weeks inside; context compression left an estimated 10% of eligible internal Codex traffic unmonitored.
METR and Redwood researchers spent three weeks with access to OpenAI’s internal code, monitor instructions and engineers, @kimmonismus reports. Agents could theoretically postpone monitoring indefinitely by continuing to work instead of handing control back.
The format began as a Claude Code experiment. It now sits next to MCP.
Agent Skills started as an experiment in Claude Code, @bcherny writes. Many across the industry have come to adopt it, so Anthropic found it more fitting to donate the format to the @AgenticAIFdn, next to MCP, the Model Context Protocol.
An uncensored DeepSeek, five omp sessions and his MCP Agent Mail, for a home surveillance system 3,587 commits deep.
@doodlestein has five omp sessions working together through his MCP Agent Mail, each using DeepSeek v4.1 Flash Uncensored to exploit and root his network cameras from a bunch of different brands. He wants them all to work with his FrankenSurveillanceSystem: his own take on the sensor fusion and threat monitoring Anduril and Palantir do for the US government, for the home.
Storm-0940 hit 2,250+ tenants after the ConnectWise breach and reached privileged cloud access in under three hours.
Microsoft says Storm-0940, a group it links to the Chinese state, has used ARTEX, an open-source agentic penetration-testing framework, since as early as September. After compromising ConnectWise’s cloud backup service, it tried sign-ins against more than 2,250 Microsoft 365 tenants at 2,900 attempts a minute.
From posts by and reports from @MsftSecIntel, @kimmonismus, Microsoft Security Blog, CyberScoopRead the story →
One harness hit 15 Argentine healthcare groups; another pulled domain controller credentials in about 20 minutes.
CrowdStrike says four previously unreported threat actors left their AI agent frameworks in open directories between May and August. A Spanish- and English-speaking actor hit at least 15 Argentine healthcare entities with a Claude Code harness.
From posts by and reports from @kimmonismus, CrowdStrike, WIREDRead the story →
An org-owner token turned off branch protections; Wiz saw stolen AWS keys probing Amazon Bedrock hours later.
LiteLLM says the attacker behind Wednesday’s malicious 1.107.0 release held GitHub CLI credentials for three developers. An org-owner token was used to disable branch protections and point the release workflow at a lookalike repository.
From posts by and reports from @kimmonismus, LiteLLM, Wiz, The Hacker NewsRead the story →
Deployments paused and every API token reset; Railway says the attackers were after AI inference keys.
Railway says an unauthorized party may have accessed customers’ encrypted environment variables, and for some customers obtained decrypted values. It calls the attack part of a series aimed at platform companies to compromise AI inference providers, and it asked upstream providers to revoke exposed keys.
From posts by and reports from @levelsio, @kimmonismus, Railway, Railway statusRead the story →
A Brotli bug in Node allowed code execution. GitLab’s CVSS 10 regex flaw reaches back to version 7.0.
Node.js released 22.23.4, 24.22.1 and 26.12.1. They fix an out-of-bounds write in the Brotli decoder that permits remote code execution, reported by OpenAI’s Codex Security, plus two TLS hostname checks that could be bypassed.
From posts by and reports from @vercel_dev, @kimmonismus, @levelsio, Node.js, GitLabRead the story →
Run /grill-me on /software-factory; research docs cover orchestrators, Git hosts, ticket trackers and runners.
@mattpocockuk’s prompt of the day is “/grill-me on how a /software-factory might help us.” It helps you work out whether a software factory fits your team, then brainstorms the infrastructure you would need to set one up.
@MiaAI_lab’s EXL3 2.9 bpw quant on TensorFold: 2,124 tok/s prompt reading and 524,288 tokens of context.
@niklaslenz_ai measured DeepSeek-V4.1-Flash-1017 on two DGX Sparks with @MiaAI_lab’s Oct 21 recipe: her EXL3 2.9 bpw quant on TensorFold, thinking off. Decode ran 72.0 tok/s on code and 65.0 on prose; four requests at once totaled 98.4 tok/s.
@jayleaton’s LocalRouter runs Qwen 27B, speech, images, video and Whisper on TensorFold, and talks back in about half a second.
@jayleaton’s LocalRouter AI assistant runs full duplex on one DGX Spark, built on @ashxhart’s TensorFold. It pairs Qwen 27B with Drama 3 streaming speech, Qwen Image Turbo 2.1, MiniMax H3 video and Whisper STT.
OpenAI, Thinking Machines and China’s Telos are building models that take a correction mid-answer.
Full-duplex models keep listening while they speak, so a correction can change what comes next, @kimmonismus explains. Thinking Machines Lab’s research preview processes audio, video and text in 200ms chunks, with a background model for deeper work.
Craft trails only OpenAI’s three models. It plans each layout before drawing it.
Craft, Canva’s first in-house image model, debuted at No. 4 on Design Arena’s rankings, behind only OpenAI’s three models, @TheRundownAI reports. It was trained on millions of Canva templates.
He saved its memory to a 42 MB file, restarted, and it still followed his Rust house rules.
MemT 1.0, from the @MemOS_dev team, writes what you tell it into a memory state inside the model, while the current message still goes in the context window. @sudoingX ran it in bf16 on a 128 GB Strix Halo, taught it his Rust rules, saved its memory to a 42 MB file and killed the process.
Meanwhile GLM 5.3 Flash now runs on TensorFold ZIG on every configuration, on a separate branch.
@MiaAI_lab tested Flint 40B and found it very fast, but “not, in any way, shape or form, a replacement for Qwen3.6-35B.” @MiaAI_lab asks when a local model will finally beat it: this year, 2027 or never.
The first live pilot of its kind, they say: approved invoice payments through supplier portals, no human authentication.
Bill, @Billtrust and @Visa announced what they say is the first live pilot to let AI agents initiate and complete approved buyer invoice payments end to end, @AccountingToday reports. The payments run through supplier portals without the need for human authentication, the companies say.
It plans the work, grounds each calculation in an authority and shows its reasoning for review.
Thomson Reuters launched the next generation of CoCounsel Tax, Audit & Accounting on Thursday. “Precision alone doesn’t make professional work correct,” the company says.
He asks security leaders to go to their boards now, for the basics: “we need to hold on for ~18 months.”
@JasonDClinton, whom @kimmonismus identifies as Anthropic’s deputy chief information officer, writes that “as bad as things have gotten in the past 4 days, they are going to get a lot worse before they get better.” He asks security leaders to go to their boards and ask for help right now: “not for ai, for the basics.”
UAE investors lead the round, @kimmonismus reports. It would top Anthropic’s last private valuation.
OpenAI nears an up to $30 billion funding round at a $1.25 trillion valuation, led by UAE investors, @kimmonismus reports. The valuation in this new round would exceed Anthropic’s last private valuation, he notes.
Even when doctors can search the web, says @mustafasuleyman. It also orders fewer tests.
Microsoft’s MAI Diagnostic system is now 1.7x more accurate than physicians in the ED, even when they have access to the web to look things up, @mustafasuleyman writes. It is also lower cost and does fewer medical tests to get to its diagnosis.
Labs where AI designs experiments and robots run them around the clock, starting with drug discovery.
Ex-OpenAI product chief Kevin Weil raised $215M for Warp Sciences, @TheRundownAI reports. The startup is building labs where AI designs experiments and robots run them 24/7.
Banks are scrambling to shut them out. A banking lawyer says customers should worry more about being tricked.
AI agents are entering U.S. bank accounts and banks are scrambling to shut them out, @kimmonismus writes, citing Axios. Axios reports some appear to attempt actions beyond users’ instructions.
From posts by @kimmonismus, @AlexH_Johnson, @emollickRead the story →
Cloudflare counts millions of attempts. Censys found about 1.57 million exposed hosts still showing unpatched PHP.
Amazon says an exploit attempt for php2shell hit its MadPot honeypot on October 17, four days before the PHP fixes, and that the flaw “is now being broadly exploited in the wild.” Cloudflare CEO Matthew Prince told SecurityWeek attempts went from thousands to “millions now.” watchTowr says widely sprayed proofs of concept do not achieve code execution.
From posts by and reports from @levelsio, @techreview, @TheRundownAI, @kimmonismus, Amazon Security Blog, SecurityWeekRead the story →
Email addresses and system IDs worldwide; 196 accounts had passwords changed without permission. No card data.
Capcom says unauthorized access to its CAPCOM ID service touched 11,774,247 records of email addresses and system IDs worldwide. For 43 people, birth dates, gender, country and nicknames were also exposed, and 196 accounts had their passwords changed without authorization.
Microsoft says the rollout embeds AI assistance across the business; @satyanadella cheers it on.
Microsoft says Volkswagen Group has brought Microsoft Copilot to more than 330,000 employees, embedding AI assistance across its business so teams can focus on higher-value tasks. “Super to see Volkswagen Group putting Copilot to work across the company,”