OpenAI Says GPT-6 Cyber Built 948 Linux Exploits

FEATURED INSIGHT💡

OpenAI reports 948 Linux exploits as GPT-6 Cyber arrives

OpenAI’s GPT-6 Cyber is reaching selected security researchers after a Linux audit produced hundreds of working exploits. In its November 4 research report, OpenAI says an automated pipeline reproduced 1,724 vulnerabilities and built 948 local privilege-escalation exploits on a commonly used Linux distribution. Those exploits could let someone who already has a low-privilege account gain root access, the system’s highest level of control. The run cost about $1.2 million in API credits.

The pipeline selected security-relevant code, investigated suspected flaws, reproduced failures in isolated environments and generated candidate patches. Human researchers reviewed the outputs before reporting them to maintainers. For software teams, the useful advance is a report backed by executable evidence and a proposed repair. A generated patch still needs review; the experiment’s counts do not establish that every affected system has been fixed.

OpenAI’s November 5 release announcement reports a 48.5% score on its ExploitGym benchmark, up from 32.1% for GPT-5.6 Cyber, and better adherence to authorized testing boundaries in its evaluations. It also reports 81 previously unknown high-severity Firefox flaws; Mozilla’s Firefox 159 advisory independently lists fixes credited to OpenAI researchers and Codex Security. GPT-6 Cyber initially reaches only selected approved Daybreak Red customers through Codex and the API. Security leads can assess the reported findings and patch workflow while checking whether their organization qualifies for access.

The voice-to-text app that built its own speech model

Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own.

It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office.

Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them.

Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing.

Try Wispr Flow free

ON THE HORIZON 🌅

METR estimates Anthropic’s AI research is advancing twice as fast

AI may already be helping accelerate the development of its successors. In a November 5 assessment, METR estimates that Anthropic is improving model capabilities on well-specified tasks at roughly twice its earlier rate. Its researchers reviewed internal data and interviewed staff between August and October. They judge that AI-assisted research likely accounts for most of the current acceleration, with substantial uncertainty; their plausible range includes about 1.3 to three times the historical pace.

The report describes progress driven mainly by AI assisting people, with fully autonomous research contributions still emerging. METR also says Anthropic redacted most of the evidence it found most informative, limiting the public’s ability to examine its conclusions. The finding raises a concrete planning question: if better models help researchers build better models faster, how quickly can evaluation and oversight keep up? The report offers evidence of a possible feedback effect within one lab, with limits that matter when applying its conclusions elsewhere.

LATEST IMPORTANT NEWS 📰

Google cuts Gemini Omni video prices 30%

Gemini Omni 1.2 Flash generates clips up to 15 seconds long, extends conversational editing and costs 30% less than Omni 1.1, Google announced November 5. The listed price is $0.07 per second for 720p video, and the model is generally available through the Gemini API and Gemini for business. Google also reports better adherence to reference images and less visual drift. For creators producing product demos or ads, longer clips and follow-up edits offer more room to revise a shot without rebuilding the entire sequence.

Meta’s Muse gets a voice for working out loud

Meta is rolling out voice conversations and expressive avatars in Muse, starting in the US across mobile, web and Mac. Users can interrupt the assistant, choose voice-only interaction or show its avatar, and talk while Muse works on background tasks. Meta suggests rehearsing interviews and presentations or planning trips through conversation. The addition gives people a way to develop an idea aloud and respond to questions as they go; voice on AI Glasses and Muse Charm remains planned rather than available in this rollout.

NVIDIA gives robots more worlds to practice in

NVIDIA detailed four new Cosmos 3 Nano checkpoints on November 5, giving developers different ways to generate driving and robotics simulations. They cover multi-view driving data, streaming camera and LiDAR observations, video rendered from simulator controls, and video predicted from robot actions. Open weights and adaptation recipes let teams work with their own sensors and environments. These are additions to the existing Cosmos 3 family, with separate roles in a simulation pipeline. NVIDIA explicitly identifies faithful policy learning and evaluation as work still requiring measurement.

Create work your clients can rely on

With Lovable, you'll have a working version of your client's site ready for the first meeting. Plus:

• Launch on your client's own domain

• Build sites that get found on Google

• Catch bugs and errors with project monitoring

• Run every client site on one plan

Start building, $5 off

^{_{Discount of $5 applies to first month of any paid plan. You need to sign up via this webpage in order to claim the discount. The discount applies to new and returning users without an active paid plan. Terms apply.}}

FOR THE TECHNICALLY INCLINED 🛠️

Inkling 2 Small packs 594 billion parameters, activates 17 billion

Thinking Machines Lab’s Inkling 2 Small, introduced November 5, uses a mixture-of-experts design: it stores 594 billion parameters but activates about 17 billion when processing a token. Its model card describes routing each token to eight of 384 experts, with a combination of local and global attention and support for image inputs. Sparse routing lets the model use selected specialist components without running every parameter for each token. The company trained it on long coding tasks across multiple agent setups, including rewards that assess how it approached the work. Developers can already sample and fine-tune it on Tinker. Downloadable weights under Apache 2.0 are planned after external safety testing; they were not available at launch. That distinction matters for teams choosing between testing through a hosted platform now and running their own model later.

AI TOOL OF THE DAY 🚀

Runway App Studio lets paid subscribers describe a repeatable image, video or audio workflow in plain language, turn it into a hosted creative app, and share it with teammates using their existing credits.

See what's really happening in every deal

Aligned shows you what buyers are doing between meetings inside your most important deals then surfaces the risks, openings, and next steps. No more guessing where a deal stands or finding out a champion went quiet too late. Just real visibility, so you can act before it's too late.

Start Using Aligned Free