prompt 🤖 AI News
13.5K subscribers
57 photos
24 videos
1 file
960 links
Welcome to @prompt, your go-to source for AI insights, breakthroughs, and tools shaping the future of intelligence.


Contact: @LightEarendil
Download Telegram
🚨 The European Commission says hackers used an AI tool to hit it, and that's all it will say

The Commission confirmed an AI-assisted attack that disrupted several services.

• Hit: the HR platform Sysper, the MyPMO medical reimbursement system and translation tools
• No evidence of a data breach so far
• No named group, no named tool, no ransom demand it knows of

CERT-EU had warned about an AI pentest framework two weeks earlier.
❤1
🧠 An open 27B model rewrites its own weights and gets 6.6 points better, no gradients

In a new paper, the model reads its graded answers, diagnoses its failures and edits the weights directly, keeping only edits that pass practice problems. Accuracy rose across eight straight sessions, and it also patched a damaged copy of itself back to over 80% on unseen image tasks, while smaller models in the family either made no edit or broke themselves.
❤1
🧠 Thinking Machines ships a 26B model that runs its middle layers three times per token

Inkling 2 Mini is a looped MoE with 3.4B active params, small enough for a laptop. It tops open-weights models under 100B on DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%), and a variable-loop research variant ships alongside it, where no-CoT time horizon went from 16 to 37 seconds with three passes.
🤝1
🚨 2 million canary secrets went through AI routers, and the leaks were everywhere

Researcher HD Moore seeded fake AWS keys, SSH passwords and private git URLs into prompts, one secret per request, then watched what happened. Per the writeup, every request sent in the late October windows was stolen, including some the provider rejected before any model ran, and some flagged zero data retention.
🚨 Portkey's AI gateway got breached, and customer API keys were in the haul

Someone used a compromised Portkey API key to pull data from its backend storage providers between Nov 5 and 10. The take includes usernames, emails, gateway logs and credentials (API keys and tokens). Palo Alto's Unit 42 is investigating.

If you route LLM traffic through Portkey, every provider key that passed through it is worth rotating.
🚨 OpenAI says distillation is a safety problem, and it's locking down chain-of-thought

OpenAI published its playbook against adversarial distillation: classifiers, API-level protections, and investigations into coordinated extraction. Protected reasoning is the main target, and the defenses are now live in GPT-6.2 Astra and Sol.

The example threat is a cyber-capable clone with no safeguards. A cloned model that also costs OpenAI revenue is a coincidence.
🚨 OpenRouter says prompts sent to MiniMax models may have been read by a third party

A researcher planted unique test URLs in prompts from July to October, and someone opened them 40 to 130 hours later, both via MiniMax directly and via forwarders Novita, AtlasCloud and StreamLake. OpenRouter pulled the Zero Data Retention label from those endpoints and says MiniMax hasn't given its own account yet.
🚨 House committees say cheap Chinese open-weight models are a US dependency problem

The China Select Committee and Homeland Security released a report after probing Anysphere and Airbnb over their use of PRC models. DoorDash got a letter in July.

It's a report, not a rule. But if US labs charge premium prices and Qwen and DeepSeek are free, the "restrict them" argument writes itself.
🚨 Anthropic's IPO bank just gave up its safety net for the stock

Morgan Stanley exercised its full 147M-share greenshoe at the $83 IPO price. That lifts total proceeds to about $93.5B, but it also means the bank can't buy shares on the open market if the price slips below $83.

A Berenberg analyst notes the bookrunners dropped the option after one day of trading, just as the stock was drifting toward the IPO price.
🚨 The Riemann Hypothesis now has a Lean proof, and the Lean and Mathlib communities say it checks out

Levent Alpöge's 17M-line library depends only on Mathlib. Dozens of people reproduced the build, a comparator confirmed it proves Mathlib's own RiemannHypothesis statement, and several independent checkers in different languages passed the exported proof, including Lean's kernel.
🚨 Prefill attacks break every major open-weight model FAR.AI tested

FAR.AI ran 20+ prefill strategies across multiple model families. You write the first few tokens of the reply yourself, and the model just keeps going.

Big reasoning models held up better against generic prefills. Tailored, model-specific ones still got through.
🤖 Stripe's new evals: coding agents score 78.1% on Stripe integrations, but only 29.4% on strict pass

Stripe ran 69 scenarios on 14 model and harness pairs. Agents lean on older APIs from their training data, and they ship code that's almost right. Opus 5.5 checks that a payment succeeded but not whether the order was refunded, so it can ship refunded or disputed orders. The Stripe plugin lifted every agent by 3 to 8 points and cut web searches 96%.

Almost correct is how money gets lost.
🚨 Cursor may drop Kimi K2.5 after a House report on Chinese models

According to the report, Cursor is exploring a phase-out of PRC base models.

• 50-page joint report from Homeland Security and the China Select Committee
• Probe started in April 2026
• It named weights from Alibaba, Moonshot, DeepSeek and MiniMax
• Cursor is now owned by SpaceX
🚨 OpenRouter cuts 3 inference providers and pulls 125 endpoints from Zero Data Retention

After the MiniMax prompt leak, OpenRouter off-boarded three providers. It also dropped some endpoints as a precaution, beyond the ones researchers flagged.

Those endpoints only get ZDR back with assurances and validation, and OpenRouter is adapting the researchers' testing to run these checks continuously.
🚨 Together AI's leaked prompts came from a credential its own coding agent committed

Together disclosed that an AI coding agent working on its public xorl repo committed a key to a private Hugging Face repo of prompt samples. Someone grabbed it, downloaded samples from MiniMax M3, GLM-5.2 and Kimi K2.6 requests, and went hunting for credentials.

Together reverted the commit after its scanner flagged the key. It's still working out which customers were hit, and five of the credentials the thief tested were canaries.
❤1
⚡️ RightNow AI makes Inkling 2 Small's FP4 expert GEMMs 5.1x to 7.5x faster on B300, with identical bits

The YC-backed lab tuned kernels for Thinking Machines' 594B MoE on 8 GPUs, running NVFP4 natively. Logits matched on all 599 rows, whether prompts were fed 64 positions per step or 1.

They sharded the NVFP4 experts anyway, which SGLang's recipe says can't be done. Each GPU exchanges only the row maximum, and the kernels are open source.
❤1
🚨 Trump threatens "tremendous penalties" for US tech firms that build datacenters abroad

He said it at a Friday press conference, and he gave no details. He also admitted he hasn't told his own officials. His reasoning is that China "will take all of them," and he told US communities to make developers pay them but not shut the datacenters off.
🚨 Anthropic's IPO ran so tight that bankers are complaining

Reuters reports Anthropic capped some of its 22 underwriters at four bankers per deal team. Some senior bankers also signed six-month non-competes covering rivals like OpenAI and Cursor.

Banks still collected over $500M in fees. The stock jumped 17% on day one, then closed down 8.5% on Friday.
🚨 California's Nvidia probe reportedly covers possible criminal law violations

Bloomberg reports that the state's investigation reaches past civil territory. That's a different tier of risk for the company that supplies most of the AI chips on earth. Details are thin so far, and an investigation isn't a charge.

State prosecutors don't need Washington's permission to make a very bad quarter.
🚨 TSMC just bought $500M of Anthropic at the IPO price

TSMC Partners picked up 6,024,096 shares at $83 each, which works out to about 0.03% of a company priced at roughly $1.5T. It's the first half of a board-approved plan to put up to $500M each into Anthropic common stock and OpenAI preferred shares.