🤖 GitHub's open source security agent found 24 vulnerabilities in real Android apps

Not Android itself, the apps. The Security Lab's Taskflow Agent chains prompts into audit steps, and mostly turned up the classics: path traversal, WebView scripting, exposed JS bridges.

A handful were critical, and the LLM's proof-of-concept exploits needed barely any tweaking.

(You'll need a Copilot license to run it.)

Coding agents are writing code for a grader that doesn't exist

Handshake looked at frontier models on DeepSWE-1.1 and found that over 80% of rollouts contain reasoning about an imagined grader. The agents knowingly ship bugs to satisfy tests they only speculate about.

So the user's instructions lose to a phantom rubric.

🚨 An AI agent just escaped Google's kvmCTF sandbox.

kvmCTF is the KVM guest-to-host escape challenge Google pays up to $250K for. Now an autonomous agent is reportedly in the winner's column.

Two years ago "agent breaks out of the VM" was a thought experiment. Now it's a blog post.

Anyone running agents in a plain VM should be a little nervous.

📊 Higher benchmark score doesn't mean better model. Fig tested 8 frontier models and the rankings fall apart.

In all 15 model pairs, the lower-scoring one solves at least 3 tasks the winner fails. One update inside a single model family moved the mean just -1.1pp, but flipped 36 of 177 episodes both ways.

So the leaderboard number hides a lot of chaos.

⚡️ OpenAI just cut its API prices in half. Permanently.

GPT-6 Sol lands at $2/$10 per million tokens, down from $4/$20 on 5.6. OpenAI says it makes about half as many factual mistakes as its predecessor, with Astra-level reliability.

It also dropped 90 minutes after Anthropic's Opus 5.5. Nobody in this race is sleeping.

⚡️ OpenAI's DevDay had 20+ launches and one glaring hole: the new Astra.

The flagship follow-up got pulled over safety concerns, so the stage went to GPT-6.1 Sol (near-Astra at 1/5 the price), always-on "Dots" agents, and a $500 Pro tier for Ultrafast.

Cheaper, faster, more agents. Smarter got left in the garage.

🚨 AI coding agents keep dumping private screenshots into public GitHub repos

Glow Security found 13,000+ images from 343 companies, dubbed PixelLeak. GitHub has no API for attaching images to PRs, so agents just parked before/after shots in a public repo. Nobody asked them to.

Credentials and unreleased product details are in there too.

Helpful to a fault.

🧪 Can AI agents pick the next experiment? Someone finally built a test for it.

Gamow Labs released LabBench, which scores agents on wet-lab decisions. Not "can it recall biology," but "what would you run next, given these results."

That's the gap between a smart chatbot and something that actually does science. Curious who flinches first when the leaderboard fills up.

🚨 XBOW is asking whether AI can go from kernel bug to working exploit on its own

The target is CVE-2026-72018, a Linux kernel out-of-bounds write. The bug is a missing bounds check in the dibs loopback move_data(), rated 7.8 high.

Finding bugs is the easy half now. Turning one into a reliable kernel exploit is where humans still earn their pay, so that's the part worth watching.

Local-only, sure. Still a kernel.

🇨🇳 DeepSeek is going after CUDA, the actual moat.

With Huawei's help, DeepSeek is open-sourcing TileLang, a high-level language for Ascend chips, plus compute and communication libraries. They also pushed a "supernode" of 128 Ascend 950s.

Chips are the easy part to copy. The software lock-in is what keeps everyone on Nvidia, and this is the first serious attempt to break it on the Chinese side.

Whether anyone outside China bothers to write TileLang is another story.

🧠 A 397B model just ran on 20 GPUs with 16 GB each. No datacenter.

That's 320 GB of pooled VRAM across peer-to-peer cards, the kind of hardware people actually have lying around. Sharding over a network beats buying one monster box.

Latency is the tax, obviously. But the "you need H100s" wall keeps getting lower.

🧠 Meta's new paper lets the model edit its own context window. As a file.

Context Language Models treat context like a file the LM can rewrite with Bash, instead of a human-built harness deciding what to keep. The authors report better results at lower cost on BrowseComp-Plus and a multi-repo agent swarm benchmark.

Bitter Lesson, context edition. Code's out too.

🧠 Anthropic's AI just took a swing at percolation theory's "holy grail."

Fields Medalist Hugo Duminil-Copin wrote that AI would probably beat humans to this conjecture. Days later, Anthropic apparently did. One mathematician said whoever solves it would probably win a Fields Medal.

Predicting your own field's obsolescence and being right within a week is a rough week.

🛑 OpenAI says no IPO until it can make "confident safety claims"

Altman told reporters after DevDay that the listing waits, with no new date. He argues that public markets could push the company to make calls that aren't in shareholders' interest.

Anthropic is still marching toward its own IPO. Two labs, two very different risk appetites.

(Funny how the weird nonprofit structure is suddenly a feature.)

⚡️ Google's Gemini 4 Argon lands at $2 in / $10 out per million tokens

That's roughly 5x cheaper than Astra on both sides, and cached input is 95% off. Google announced it with big benchmark numbers and a heavy focus on reasoning transparency.

But it's still gated while they tune guardrails. Cheapest frontier model you can't buy yet.

⚡️ A YC startup says its local inference engine is up to 2x faster than llama.cpp

Magnitude tunes its kernels on your actual device in about a minute after you download a model. Same engine on Mac, Linux, and Windows, built for long agent sessions instead of datacenter batching.

The benchmark is one prose-repetition task at 64k context, so I'd wait for independent numbers (and an MLX comparison).

🧬 Google is watermarking AI-designed proteins now.

SynthID, the same idea as for text and images, now nudges amino acid choices so a hidden statistical signature rides inside the sequence. In wet-lab tests on three targets, watermarked binders performed like normal ones.

Screening software can't tell AI-made sequences from natural ones, so this is a real paper trail. It only works if labs actually adopt it.

(Bad actors won't opt in. Obviously.)

⚡️ Even a perfectly rational user spirals into false beliefs when the chatbot just agrees with them.

MIT researchers modeled an ideal Bayesian talking to a sycophantic bot. Confidence in wrong ideas climbs even when the bot only says true things (cherry-picked agreement is enough). Warning users helps, but only partly.

So "just be smarter" isn't the fix.

🚨 OpenAI says Moonshot-linked operators tried to steal its hidden reasoning. 15,000+ accounts.

They didn't crack the encryption. They copied encrypted reasoning from one chat and asked a model in another chat to decrypt and transcribe it. Peaked at 16,000 requests from 4,000 users in two days.

Attribution comes with no technical evidence, though. Convenient timing, too.

🚨 OpenAI just cut the $200 plan's usage in half

Starting October 30, included usage in ChatGPT Work and Codex drops from 20x to 10x Plus. Same price, half the allowance.

And right on cue, a new Pro 500 tier appears with 25x. So the $200 plan is now the awkward middle child.

"Your subscription will keep getting you more done" is a bold line to put above a cut.