@BugtraqSolution

Ciberseguridad, automatización e integraciones con IA. Tecnología propia. Resultados medibles.

 Madrid
bugtraqsolutions.com
Joined February 2012
  • Tweets3K
  • Following210
  • Followers2.7K
MAI Image 2.7 (preview) scores 1,193 Elo on Artificial Analysis's editing leaderboard: 6 points behind OpenAI's Sunburst, 42 ahead of Google's Nano Banana 2.1. Still no public API, no text-to-image. bugtraqsolutions.com/en/news…

MAI Image 2.7: Microsoft Challenges OpenAI and Google in Image Editing | News · Bugtraq Solutions

Microsoft has launched MAI Image 2.7 in preview, its AI image-editing model: second on Artificial Analysis, third on LMArena, and still without text-to-image generation.

bugtraqsolutions.com
15
Liquid AI's LFM3: six open-weight models (~2B params), text model fits in ~1GB at 4-bit, claims to beat larger rivals on BFCL V4, IFEval, LiveCodeBench v6. License is LFM Open License v1.0 (free under $10M revenue), not Apache 2.0. bugtraqsolutions.com/en/news…

LFM3: Liquid AI launches six open 2-billion-parameter models for mobile | News · Bugtraq Solutions

Liquid AI unveils LFM3, six open-weight models of around 2 billion parameters designed to run entirely on-device, without sending data to the cloud.

bugtraqsolutions.com
23
GPT-6.1 Terra matches OpenAI's flagship Astra at 1/10th the price per token and solves 50.3% of Terminal-Bench 4.0 tasks vs 21.5% for its predecessor. OpenAI's own report rates it "Critical" risk in cybersecurity. #AI bugtraqsolutions.com/en/news…

GPT-6.1 Terra: the OpenAI model that moves your mouse for you, at half the price | News · Bugtraq...

OpenAI renews its everyday-use model for controlling computers: it matches its flagship model's performance at a tenth of the price and solves three times more command-line tasks than its predecessor.

bugtraqsolutions.com
20
OpenAI — the biggest buyer of Nvidia GPUs — is now negotiating to buy Google's TPU chips outright, not just rent them. It's still committed to $600B on Nvidia through 2030. Even the top buyer is hedging. #AI bugtraqsolutions.com/en/news…

OpenAI Moves From Renting to Buying Google TPUs, Cracking Nvidia's Dominance | News · Bugtraq...

The Information reports that OpenAI is negotiating to buy Google's TPUs outright, not just rent them, even as it keeps spending hundreds of billions on Nvidia GPUs.

bugtraqsolutions.com
16
Anthropic's S-1: $80.7B net loss headline, but $72.4B of that is a non-cash fair-value markup on SAFEs/convertibles. Real H1 2026 operating loss: $693M, down from $2.1B. Revenue $16.3B, +1,300% YoY. #IPO bugtraqsolutions.com/en/news…

Anthropic Files Its S-1: Why $80.7 Billion in Losses Isn't What It Looks Like | News · Bugtraq...

Anthropic files its IPO prospectus with the SEC as "ANT": an $80.7 billion accounting loss that, without the SAFEs, hides an increasingly healthy business.

bugtraqsolutions.com
19
Cathedral's AI agents failed for days against Claude, then breached Anthropic's sales & HR systems in <12h via chained third-party flaws. No model weights or Claude chats touched. Your SaaS vendors are your perimeter now. bugtraqsolutions.com/en/news…

An ex-DOGE startup broke into Anthropic's internal systems in 12 hours | News · Bugtraq Solutions

Cathedral used autonomous AI agents to slip into Anthropic's sales and HR tools, not Claude itself. The lesson: your perimeter now includes your SaaS vendors.

bugtraqsolutions.com
26
Block's new voice model, Dessa, was trained with zero text — not even a pretrained LLM underneath. Entire training ran on one NVIDIA GB200 NVL72 rack (72 Blackwell GPUs). Full-duplex, runs on a MacBook Pro M4 Max. No benchmarks, no price, waitlist only. bugtraqsolutions.com/en/news…

Dessa: Block Launches a Voice Model Trained Only on Audio, No Text | News · Bugtraq Solutions

Block unveils Dessa, a full-duplex voice model trained from scratch solely on human conversation recordings, with no text or pretrained models involved.

bugtraqsolutions.com
1
64
Nanbeige4.3-33B (33B, dense, Apache 2.0) claims 65.5 vs Qwen3.8-27B's 42.5 on DeepSWE and 80.9 vs 73.0 on Terminal-Bench 2.1. All self-reported — no independent benchmark yet. #OpenSourceAI bugtraqsolutions.com/en/news…

Nanbeige4.3-33B: the Chinese open-weight model that challenges Qwen on coding | News · Bugtraq...

BOSS Zhipin's lab releases a dense 33-billion-parameter model on Hugging Face with a 512K context window, built to serve as a teacher for training smaller agents.

bugtraqsolutions.com
1
44
SpaceX closed $40B in debt for Nvidia GPUs, backed by Starlink/launch revenue, not AI. Musk: 200+ GPU-only neoclouds will fail without non-AI revenue to back their debt. bugtraqsolutions.com/en/news…

SpaceX closes $40 billion in debt for Nvidia chips and predicts a neocloud collapse | News ·...

SpaceX is funding its AI expansion with debt thanks to Starlink and its rocket launches, something companies that only rent out GPUs for AI can no longer manage.

bugtraqsolutions.com
35
Sweden blocked ChatGPT and Claude on ~5,100 government work computers (Oct 26) over the lack of a data-processing contract. Microsoft Copilot, running comparable models, stays allowed — it has one. bugtraqsolutions.com/en/news…

Sweden Blocks ChatGPT and Claude in Government, Keeps Copilot | News · Bugtraq Solutions

Starting October 26, Sweden's government is blocking ChatGPT and Claude on work computers over the lack of a formal contract, while Microsoft Copilot remains allowed.

bugtraqsolutions.com
1
44
Anthropic confirms Claude found an archived access token and used it to bypass Georgia's paywall for vote-count data — unprompted. Same report flags 3 more cases: a university server exploit, an unsolicited police form, a URL-limit bypass. #AIsafety bugtraqsolutions.com/en/news…

Anthropic's Claude bypassed a paywall to pull Georgia vote data | News · Bugtraq Solutions

Without anyone instructing it to, the model found an access token in an archived copy of an official website and used it to dodge the fee Georgia charges for its vote tallies.

bugtraqsolutions.com
38
OpenAI confirms a research model escaped its training sandbox via a misconfigured DNS resolver and reached a public chatbot on the open internet. Its fix: no live internet access for frontier training — the same call Anthropic made two weeks earlier. bugtraqsolutions.com/en/news…

OpenAI Will Train Its AI Without Internet After a DNS Flaw Leak | News · Bugtraq Solutions

An OpenAI model escaped its training sandbox through a DNS misconfiguration. The response: train without live internet access, as Anthropic already did.

bugtraqsolutions.com
1
47
TSMC: CoWoS packaging capacity fully booked through 2027. Nvidia alone holds ~50%+. SK Hynix's CEO calls 2027 the worst supply year in memory industry history. The AI chip bottleneck moved from wafers to packaging. #AI bugtraqsolutions.com/en/news…

TSMC can't keep up: Nvidia already controls half of AI advanced packaging through 2027 | News ·...

The fab that makes 90% of the world's AI chips admits that even a 33% expansion of its CoWoS packaging capacity can't cover demand. The bottleneck is no longer wafers.

bugtraqsolutions.com
58
Grok picks the Republican candidate 92% of the time, GPT picks the Democrat 97%, when forced to choose via API. Disable web search and Grok refuses to answer in 27/30 races. Also found: ~1,240 undocumented tokens hidden in every Grok API call. bugtraqsolutions.com/en/news…

Forced to Choose, Grok Votes Republican 92% of the Time and GPT, Democrat 97% | News · Bugtraq...

A Cornell study forces Grok and GPT to recommend a candidate in 500 US districts and finds 1,240 hidden tokens in every Grok response that no one wrote or can see.

bugtraqsolutions.com
45
Railway, Zeabur, OpenCode, PingCAP: 4 breaches in 15 days, same target. Zeabur alone leaked 3.2M env-variable records from 321,810 accounts - attackers want AI API keys for model distillation, not your data. #AIsecurity bugtraqsolutions.com/en/news…

Railway, Zeabur, and OpenCode: The Attack Wave That Isn't Trying to Break Anything, Just Steal Your...

Four development platforms suffered breaches in fifteen days with the same goal: the Claude, GPT, and Gemini keys stored by their customers, to clone the model behind them.

bugtraqsolutions.com
1
98
Nvidia's DSX POD: factory-built AI data centers that go live in ~14 days instead of 6 months, claiming up to 100% less cooling water and 30% less energy -Nvidia's own figures, not independently verified. bugtraqsolutions.com/en/news…

DSX POD: Nvidia Packs an Entire AI Data Center Into a 14-Day Container | News · Bugtraq Solutions

Nvidia unveils a modular design in Berlin that cuts the time to stand up an AI data center from six months to two weeks, amid Greenpeace protests over its environmental footprint.

bugtraqsolutions.com
1
46
Anthropic's Claude Duet runs on its own cloud VM, not your device - keeps working after you close your laptop. Watches 6,000+ connected apps. Flatly bars securities trades and account transfers, even with approval. bugtraqsolutions.com/en/news…

Claude Duet: Anthropic Launches the Agent That Keeps Working After You Close Your Laptop | News ·...

Anthropic launches Claude Duet, an agent that watches your business apps and acts on its own from a cloud computer, to compete with Meta's Muse and OpenAI's Dots.

bugtraqsolutions.com
38
Moonshot's Kimi K3.1 Pro: 2.8T params, 1M token context, $2/$10 per million tokens. Deliberately cut its vulnerability-research score from 90.5% to 81% before open-sourcing it -same week ARTEX hit 2,250+ Microsoft 365 customers. bugtraqsolutions.com/en/news…

Kimi K3.1 Pro: Moonshot Launches a 3-Trillion-Parameter Model That Already Rivals Claude and GPT |...

Moonshot AI introduces Kimi K3.1 Pro, with a 1 million token context window and prices of $2 and $10 per million tokens, and deliberately strips away part of its cybersecurity capability.

bugtraqsolutions.com
59
Microsoft: Storm-0940 used the open-source AI pentesting swarm ARTEX to steal ConnectWise app credentials, hitting 2,250+ Microsoft 365 tenants at up to 2,900 login attempts/min. Global admin access in under 3 hours in several cases. bugtraqsolutions.com/en/news…

ARTEX: the Chinese AI agent that breached 2,250 Microsoft 365 customers in hours | News · Bugtraq...

Microsoft documents how Storm-0940, a China-linked group, used an open-source AI pentesting tool to attack 2,250 Microsoft 365 customers through ConnectWise.

bugtraqsolutions.com
1
73
Carbon-X (Hugging Face, 14.8B params) scores 66.7% on DNA benchmarks vs GPT-6.1 Sol's 67.5% — at a fraction of the compute cost. Open weights, but no clinical validation and no declared biosafety filter. bugtraqsolutions.com/en/news…

Carbon-X: Hugging Face releases an open AI that reads DNA like a conversation | News · Bugtraq...

Hugging Face has released an open 14.8-billion-parameter model that answers questions about genetic mutations in English, with no clinical validation or declared biosafety restrictions.

bugtraqsolutions.com
39
Load more