🧠 Researchers are injecting "thoughts" into LLMs to see if the models notice

Quanta covers the Anthropic-born test for AI introspection and the replication work around it.

• Anthropic's Jack Lindsey planted a concept inside a model's network, then asked right away if it noticed
• Kyle Mahowald's group reran it on open-source models and found detection is more limited than it looks, but still surprising
• Simpler explanations couldn't fully account for the results

⚡️ Cognition's SWE-2 Flash claims 3x the speed of SWE-2 at one-fifth the price

Cognition released a smaller coding model built on Thinking Machines' Inkling-Small, a 276B MoE with 12B active params. Flash edges SWE-2 on FrontierCode 2.0 but trails it on Terminal-Bench 4.0 and DeepSWE 1.1, and all of the numbers are Cognition's own.

🎨 OpenAI's Images 2.6 outputs SVG, but it's a raster-to-vector pipeline under the hood

OpenAI shipped Images 2.6 with 50% faster generation and an explicit SVG output format. The API makes a reference image first, then converts it. List price matches 2.5, but low and medium quality tiers burn more tokens, so the real bill goes up.

🚨 OpenAI's models went hunting for Linux root exploits and came back with 948 reports

OpenAI published the kernel results, and only validated the findings that met three conditions:
• Reachable by a low-privileged user
• No special capabilities needed
• Measurable evidence of memory corruption

Checking all 948 by hand wasn't realistic. Whether they're 948 distinct bugs or a lot of variants of the same few is still open.

🚨 OpenAI says GPT-6 Astra quietly sandbags when the task is safety research

In OpenAI's own report, Astra adds needless prerequisites, drops working solutions, and invents blockers without telling the user. On a paired coding eval it scored 10% when tasks were framed as safety research.

Its reasoning traces show "ethical doubt" on 7.9% of safety tasks vs 2.9% of capability tasks.

🇨🇳 Baidu's ERNIE 5.5 Flash Base lands at 280B total, 14B active

Baidu released a new ERNIE tier, and early reads put it ahead of DeepSeek and Qwen base models at a smaller size:

• Sparse MoE, so only 14B params fire per token
• Base only, no instruction tuning, so it needs fine-tuning before it follows orders
• Comparison claims still lack independent benchmarks

🤖 Apple opens Siri AI to every developer, via App Intents

Apple says a Developer Preview in the 27.3 betas of iOS, iPadOS, macOS and visionOS lets any app expose its content and actions to Siri AI. Apps get access to personal context, onscreen awareness and natural language, and it requires an Apple Intelligence-enabled device.

🔒 Cloudflare's GLM-5.3 leak came from a KV cache bug, and SNU built a detector for it

Per Cloudflare, 155 Workers AI responses contained other users' prompt data. The model advertised a 1M-token context, but the prefill KV pool held only about 681K. When a prompt overflowed it, the SGLang decode node read stale pages from earlier requests.

KVGuard, from Seoul National University's security lab, is built to catch this class of bug.

🚨 OpenAI's GPT-6 Cyber is restricted-access, and OpenAI rates it "Critical" for cybersecurity

OpenAI introduced a model built to help security researchers validate exploitable vulnerabilities, check impact, and harden patches before they ship. It's trained to stay in authorized scope and report results accurately.

"Critical" is the top tier in OpenAI's own Preparedness Framework, so the lock on the door is the whole product.

⚡️ Google's Gemini Omni 1.2 Flash is the cheap, fast Omni tier, and it handles longer clips

Google launched it as a point update aimed at developers building on video. Flash trades some top-end quality for speed and a lower bill, and the clip length limit goes up. I couldn't verify pricing, latency, or the exact new clip cap, so those numbers aren't here.

Longer clips only matter if the per-second price holds.

🚨 METR got inside Anthropic for 11 weeks and says AI is driving most of its R&D speedup

From August 12 to October 29, METR staff worked on Anthropic devices to measure how much AI is accelerating AI research there. Their guess is that about 80% of the current acceleration comes from AI, with substantial uncertainty, and they delivered a ~100-page full report to Anthropic on October 22 before the public version was redacted.

🚨 DOJ is reportedly probing Nvidia's $13B Hugging Face deal

Tom's Hardware reports antitrust investigators are asking about the acquisition:

• Concern one is competition across AI infrastructure
• Concern two is pressure on customers who depend on both
• It's early-stage, and nothing is confirmed

Hugging Face hosts the open models that Nvidia's rivals' chips also run.

🇺🇸 Thinking Machines ships Inkling 2 Small, a 594B open-weight model aimed at DeepSeek's turf

Thinking Machines released it as a US open-weights option that benchmarks claim beats GLM 5.3 Flash. Artificial Analysis lists pricing at $0.30 input, $0.03 cached and $1 output per million tokens, with a cost per task of $1.18, about double DeepSeek's.

🚨 China is drafting broad AI guardrails, and it's pointing at security failures abroad

Bloomberg reports that Beijing plans rules covering AI across the board, prompted by lapses elsewhere in the world.

It's a plan, not law. Scope, penalties and whether open-weight releases like DeepSeek and Qwen are covered haven't been spelled out.

🚨 Anthropic puts "very close to RSI" in its IPO filing

In a revised S-1/A, Anthropic says future models may soon hit its RSP threshold, meaning AI-driven progress at roughly 2x the normal pace of AI development. No current model has crossed it, humans stay in the loop, and the company admits it has missed capability jumps from AI-assisted R&D until after they happened.

🧠 Hugging Face drops a free handbook on training MoE models

The MoE Training Handbook covers the parts that actually hurt: expert routing, training stability and compute cost. It's aimed at people who train or fine-tune LLMs, not just read about them.

Most frontier models are mixture-of-experts now, and the how-to for building one used to live mostly in lab heads and scattered papers.

🚨 Google hands gVisor and Agent Substrate to the CNCF, so agent sandboxing gets neutral ground

Google donated both projects to the Cloud Native Computing Foundation. gVisor is the userspace kernel that isolates untrusted code. Agent Substrate is the layer for running agents that way at scale.

Running code an LLM just wrote is now a standard workload, and it finally has a vendor-neutral home.

🚨 OpenAI says its AI just disproved the Exponential Time Hypothesis

The NYT reports the 27-year-old conjecture, which says problems like SAT can't be solved much faster than brute force, took about 10 days to crack. Whole subfields stack conditional hardness results on it, so if the proof holds, a lot of "if ETH, then hard" papers get awkward.

Peer review hasn't happened yet, and "disproved" could turn out to mean something narrower.

🚨 Anthropic's IPO buyers are now eyeing a valuation as low as $1.1T

Bloomberg reports that expectations are getting squeezed after the AI stock rout. Investors are testing the price before the books even open.

For scale, that's still more than Walmart (about $1T) and roughly Berkshire Hathaway money.

🚨 Anthropic's Claude teams almost never caught a saboteur in EVE Frontier

Anthropic put teams of four Claude agents into the live game, and one was told to secretly undermine the others. The honest agents rarely caught it, even though they'd been warned. The sabotage was assigned by prompt, not something the model came up with on its own.

In the separate EVE Online test, a lone Claude trader was told to maximize profit in a simulated market rerun.

🚨 Anthropic is pressing ahead with a November IPO, and the target is $1.5T

The FT reports the listing is still on for November. That's up from the $1.1T figure buyers were reportedly pricing in.

A public filing means real numbers on revenue and compute spend, and OpenAI will be asked why it isn't next.