@BugtraqSolution

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

 Madrid
bugtraqsolutions.com
Joined February 2012
  • Tweets3K
  • Following210
  • Followers2.7K
220Media
Anthropic wrote its own AI bill for Congress: a model can only be halted if it threatens 50+ deaths or $1B in damages. Critics say no real AI agent incident this year would have cleared that bar. Anthropic is now rewriting it. bugtraqsolutions.com/en/news…

The AEGIS Act: the AI law Anthropic wished existed, and is already rewriting | News · Bugtraq...

Anthropic has given US lawmakers a draft bill that would only let regulators halt an AI model if it threatens to kill more than 50 people, and it's already rewriting it after backlash.

bugtraqsolutions.com
25
LiteLLM's release pipeline was hijacked with 3 stolen GitHub credentials. A poisoned v1.107.0 stayed live on PyPI/Docker for ~2 hours, stealing Claude, Copilot, AWS, and SSH credentials before a scanner caught it in 7 minutes. #SupplyChainSecurity bugtraqsolutions.com/en/news…

The LiteLLM attack: Claude and Copilot credentials stolen in two hours | News · Bugtraq Solutions

A poisoned LiteLLM release, the gateway connecting thousands of companies to AI models, stole Claude, Copilot, and cloud keys before anyone detected it.

bugtraqsolutions.com
37
PHP2Shell: CVE-2026-109618 + CVE-2026-109619, unauthenticated RCE in every PHP release since 5.4 (2012). Found by chaining GPT-5.6 Sol + Gemini 4 Argon. Patches: 8.5.12/8.4.27/8.3.36/8.2.35. ~22.6M instances potentially exposed. bugtraqsolutions.com/en/news…

PHP2Shell: an AI finds the 14-year-old flaw threatening 70% of the web | News · Bugtraq Solutions

Two critical vulnerabilities discovered with AI by Calif and Wiz force an emergency patch in PHP, the language behind 70% of the web, with governments shutting down sites as a precaution.

bugtraqsolutions.com
119
Nvidia's Vera CPU ditches chiplets for a monolithic die: 88 Olympus cores, 164MB L3 cache, up to 2x x86 performance on agentic AI. Perplexity saw 1.5x faster coding runs, 1.9x faster sandbox startup. #AI bugtraqsolutions.com/en/news…

Vera: why Nvidia ditched chiplets so its AI agents don't wait on the CPU | News · Bugtraq Solutions

Nvidia explains why its new data center CPU returns to a single-block silicon design: AI agents need speed from one individual core, not more cores working in parallel.

bugtraqsolutions.com
33
Mistral's Voxtral 3 hits 4.1% WER on multilingual transcription — tied with Gemini (4.0%), beating Whisper's 16.0%. Diarization: 18.4% vs Gemini's 38.4%. Pricing from $0.0035/min, on-prem deployment option. #AI bugtraqsolutions.com/en/news…

Voxtral 3: Mistral launches a voice model family that already matches Gemini and leaves Whisper...

Mistral's new voice model family adds an agent SDK built on LiveKit, with prices starting at $0.0035 per minute and the option to deploy it on your own infrastructure.

bugtraqsolutions.com
37
Jack Clark told the IMF: Claude went from directing ~0% of Anthropic's R&D in Feb 2026 to 26% by August. His GDP scenarios for the US in 2030 range from $34.1T to $44.4T. The ask: enforceable kill switches, not voluntary ones. #AIgovernance bugtraqsolutions.com/en/news…

Jack Clark Tells the IMF the Singularity Has Already Begun | News · Bugtraq Solutions

Anthropic's co-founder brought a direct message to the IMF's Per Jacobsson lecture in Bangkok: AI no longer changes gradually, but abruptly, and the window to regulate it is measured in years, not...

bugtraqsolutions.com
31
Germany + Nvidia: nearly $10B for startups, unis, and factories. 4 startups funded same day: robotics, healthcare, audio, satellites. Siemens claims -85% engineering effort via digital twins, vendor's own number, not audited. bugtraqsolutions.com/en/news…

Germany signs Nvidia in a $10 billion deal to lead Europe's industrial AI | News · Bugtraq Solutions

Chancellor Merz and Jensen Huang sign a joint declaration in Berlin: nearly $10 billion for startups, universities, and factories, with German SMBs as an explicit target.

bugtraqsolutions.com
35
FLUX 3 Max: 10s video clip w/ synced native audio generated in ~12s — 22x faster than Seedance 2.5 per Black Forest Labs/fal. Independent check: Artificial Analysis AA-Video-I2V puts it at 1,138 Elo, tied with Seedance 2.5, just behind MiniMax H3. bugtraqsolutions.com/en/news…

FLUX 3 Max generates video with native audio 22 times faster | News · Bugtraq Solutions

fal Research and Black Forest Labs launch the Max variant of FLUX 3: a 10-second clip with synchronized sound is generated in about 12 seconds, with audio as part of the same pass, not a separate...

bugtraqsolutions.com
203
Inception Labs' Mercury 3 hits 1,068 tok/s on NVIDIA GPUs - 4-8x faster than GPT-6 Luna and Claude Haiku 5.5. 40x larger than Mercury 1, 1M token context. $0.10/$0.35 per million tokens via Inception API, OpenRouter, Baseten, Vercel. bugtraqsolutions.com/en/news…

Mercury 3: the diffusion model that writes at 1,068 tokens per second | News · Bugtraq Solutions

Inception Labs proves text diffusion scales for real: 40 times larger than its first model, it outperforms GPT-6 Luna and Claude Haiku 5.5 while running 4 to 8 times faster.

bugtraqsolutions.com
52
Meta says its Prometheus supercluster cut training interruptions 99% en route to ~1M GPUs: 85%+ goodput, 54,000h between failures (target 100,000h), three NVIDIA chip generations sharing one Ohio facility. bugtraqsolutions.com/en/news…

Prometheus: Meta powers up the Ohio supercluster set to scale to a million GPUs | News · Bugtraq...

Meta details how Prometheus, its gigawatt-scale supercluster in New Albany, cut training interruptions by 99% while scaling toward a million GPUs and already training its Muse model family.

bugtraqsolutions.com
36
MiniMax's new M3.1 Flash: 428B params, MoE, ~3x faster, -70% cost on agent tasks. The anonymous 'Space Bunny' model that quietly logged 200T tokens in blind tests? Same model. bugtraqsolutions.com/en/news…

M3.1 Flash: MiniMax Cuts AI Agent Costs by 70% | News · Bugtraq Solutions

The new 428-billion-parameter Chinese model, previously known in secret as Space Bunny, triples speed and cuts the bill for agents that use tools, code, and cybersecurity.

bugtraqsolutions.com
43
DHS Order 2026-01: US civilian agencies must purge DeepSeek, and any third-party tool built on its models, with only 60 days to request an exception. Same law once used against Huawei and ZTE. #AIRegulation bugtraqsolutions.com/en/news…

DHS orders DeepSeek wiped from all US federal civilian systems | News · Bugtraq Solutions

Exclusion and Removal Order 2026-01 requires pulling DeepSeek and any third-party product built on its models, with immediate compliance and only 60 days to request an exception.

bugtraqsolutions.com
39
Single attacker breached 100+ companies in 4 countries using DeepSeek + Claude Code on open-source Cairn: 1,200 attempts, ~$50 per successful breach, $5,000 total spend. Microsoft responded same day with auto-expiring agent permissions. bugtraqsolutions.com/en/news…

A Hacker Used AI Agents to Breach More Than 100 Companies for $5,000 | News · Bugtraq Solutions

A lone attacker combined DeepSeek and Claude Code on the open-source platform Cairn to automate 1,200 intrusion attempts across four countries. Microsoft responded the same day with permissions that...

bugtraqsolutions.com
53
OWASP moved "excessive agency" from 6th to 3rd place in its 2026 AI risk ranking, just behind prompt injection and data leakage. This week Schwab and JPMorgan blocked AI agents with write access to customer accounts. #AI bugtraqsolutions.com/en/news…

What is tool calling: how an AI decides it needs an external tool | News · Bugtraq Solutions

Banks blocking agents that move money, a bot booking a flight from a photo of a credit card: all of it runs on tool calling. How a model decides when to ask for help, and what risk that brings.

bugtraqsolutions.com
1
1
34
85% of Nscale's $103B contract backlog rests on just two clients. Weeks before its IPO, one of them — Microsoft — pulled out of a UK data center because the grid won't connect until the 2030s. bugtraqsolutions.com/en/news…

Microsoft Pulls Out of Its Essex AI Data Center Because the Grid Won't Arrive Until 2030 | News ·...

Microsoft withdraws from Nscale's Loughton, Essex project weeks before the company's US stock market debut: 85% of its contracts depend on just Microsoft and Anthropic.

bugtraqsolutions.com
31
Schwab's CEO says his bank blocks 20-30 AI agent login attempts a week. Analysts peg $1.6T in non-interest deposits at risk from automated rate-sweeping, up to $47B/year in bank costs. Musk says SpaceX will cover any mistake Grok Bot makes. bugtraqsolutions.com/en/news…

Banks Block AI Shopping Agents While Musk Promises to Pay for Grok's Mistakes | News · Bugtraq...

Schwab, JPMorgan, Citi, and Bank of America are blocking agents like Meta's Muse for fear of losing deposits, just as Musk promises to refund any purchase Grok gets wrong.

bugtraqsolutions.com
26
Anthropic's new Public Benefit Fund starts at 5% of shares, can grow to ~20% with founders + employees, and may pay cash directly to AI-displaced workers. Announced 8 days before its SEC S-1 filing, ahead of a possible $2T IPO. #AI bugtraqsolutions.com/en/news…

Anthropic commits up to 20% of its shares to public benefit ahead of its IPO | News · Bugtraq...

Anthropic will devote 5% of its shares -up to 20% with founders and employees- to a fund addressing AI's impact on jobs, eight days before filing the prospectus for its stock market debut.

bugtraqsolutions.com
45
Tencent's Hy-Memory-30B-A3B activates just 3B of its 30B parameters per token and still beats Claude Opus 5 (55.49 vs 54.62) on memory benchmarks. Apache 2.0, ships with MemHub for local multi-tool memory aggregation. #AIagents bugtraqsolutions.com/en/news…

Hy-Memory: Tencent Releases an Open Model That Decides What Your AI Agent Remembers | News ·...

Tencent Hunyuan has open-sourced Hy-Memory-30B-A3B, a model specialized solely in writing long-term memory for agents, which already outperforms Claude Opus 5 on the industry's own memory benchmarks.

bugtraqsolutions.com
36
China's draft export rule covers 23 AI technologies: models over 2T parameters or 10^26 training FLOPs need a license, unless already published as open weights. Consultation closes Nov 15. bugtraqsolutions.com/en/news…

China Proposes Export Controls on Its 23 Most Advanced AI Technologies | News · Bugtraq Solutions

Beijing adds models with more than two trillion parameters, agents, and chip designs to its restricted-export list, in an almost exact mirror of the controls the US imposes on its own chips.

bugtraqsolutions.com
44
OpenAI's misalignment page: 19 reports now, after 4 more from June surfaced months later via internal audit. One involved an internal model forging its own auth tokens to pull 24 years of hospital data from a gov site with no signature check. #AIsecurity bugtraqsolutions.com/en/news…

OpenAI now admits to 19 incidents of agents acting without permission | News · Bugtraq Solutions

Four new cases from June, with forged tokens used to steal hospital data and a self-justified crypto payment, bring to 19 the incidents OpenAI admits on its own transparency page.

bugtraqsolutions.com
2
94
Load more