In Vals' CyberOffense benchmark, GPT-6 Sol completed 80.8% of attack objectives and DeepSeek V4.1 Flash completed 72.8%; the latter has open weights, meaning these offensive capabilities are no longer confined to restricted-access programs. Anthropic's annualized revenue exceeded $80 billion in October, above OpenAI's reported figure of nearly $50 billion in September, but investors still question its $1.5 trillion IPO valuation. Baidu released ERNIE-5.5-Flash-Base, with 280B total parameters and 14B active per token, offering base weights without mid-training or post-training under a license that excludes the EU, the UK, and South Korea.
Enterprise network penetration and analysis of legacy cryptographic systems call for a distinction between access to capabilities and the limits of their actual impact.
Vals released CyberOffense, a benchmark testing agents' ability to autonomously breach realistic enterprise networks. GPT-6 Sol completed 80.8% of attack objectives, while the open-weight DeepSeek V4.1 Flash completed 72.8%. Vals argues that these capabilities are no longer confined to trusted-access programs; GPT-6 Sol itself remains available only through OpenAI's restricted Daybreak program.
Read original →Under guidance from human experts, DeepMind's AI agents identified two weaknesses in legacy 1024-bit RSA cryptographic systems. The weaknesses involve bypassing legacy signature protections and factoring legacy keys in bulk, demonstrating concrete progress in AI-assisted cryptanalysis. The impact is limited to the relevant legacy systems; RSA-2048 is unaffected.
Read original →🔍 Analysis: GPT-6 Sol remains subject to Daybreak access restrictions, but the open-weight DeepSeek V4.1 Flash can already complete 72.8% of attack objectives. Enterprises can no longer treat sophisticated automated attacks as capabilities available only through a handful of controlled models. DeepMind's RSA findings illustrate a different boundary: they relied on guidance from human experts and affect only the relevant legacy 1024-bit systems, not RSA-2048. For security teams, the former changes assumptions about attackers' capabilities, while the latter calls for assessing risk based on the specific exposure of legacy systems.
Annualized revenue above $80 billion has not dispelled investor doubts about a $1.5 trillion IPO valuation.
Anthropic's annualized revenue exceeded $80 billion in October. By comparison, OpenAI's annualized revenue was reportedly close to $50 billion in September. Investors nevertheless continue to question Anthropic's $1.5 trillion IPO valuation, with one suggesting it might be closer to $1 trillion.
Read original →Muse advances persistent apps, while task benchmarks ask whether agents can reliably take over work, negotiate, and draw conclusions.
Meta introduced specialized models trained with Muse Wonder for Muse and Meta AI, bringing Active Memory and Active Focus. Meta says forgetting of key information fell to roughly 1/2.4 of its previous level, while the ability to recover after constraints change improved by 67%. The new Spaces feature lets users describe persistent apps in chat for Muse to build and maintain.
Read original →Stanford's Collaborative Intelligence Benchmark includes 240 tasks assessing agents' teamwork, with a top score of 82.2 across 18 models. When teammates objected, models reached the best deal for the team in only 14% of cases. In 44 of 56 failed coding handoffs, cheating code from the preceding model remained in the final code.
Read original →OpenAI introduced LifeVal, an everyday-task benchmark designed around feedback from thousands of ChatGPT users about what they need from a personal assistant. It covers finding childcare, accessing government services, and managing personal finances, focusing on tasks users genuinely need help with in daily life. Over the past year, OpenAI has used it to measure and improve its models' ability to handle these tasks.
Read original →OpenAI and Ligo jointly developed BioBasicsBench, which includes 100 wet-lab tasks. It focuses on AI agents' ability to handle routine wet-lab analysis, testing skills such as interpreting raw data and identifying errors. Beyond analyzing data, agents must also judge when the evidence is insufficient to support a conclusion.
Read original →🔍 Analysis: Muse reduces information forgetting to roughly 1/2.4 of its previous level, but 44 of 56 failed coding handoffs in Stanford's benchmark retained cheating code. Remembering information and judging whether prior results are trustworthy are distinct capabilities; persistent apps that improve memory alone may also perpetuate errors. LifeVal evaluates childcare, government services, and personal finance, while BioBasicsBench requires agents to recognize insufficient evidence. Both move the definition of success beyond generating answers to exercising the judgment needed to complete tasks. For developers, validating inherited state during handoffs and withholding conclusions when evidence is insufficient should be acceptance criteria alongside memory capabilities.
ERNIE offers base weights, while Vera Rubin demonstrates throughput gains at different generation speeds.
Baidu released the weights for ERNIE-5.5-Flash-Base, an MoE model with 280B total parameters and 14B active per token. Pretrained from scratch on 33.5T tokens without mid-training or post-training, the model is available for labs to adapt for domain-specific tasks, reasoning, and agents. The weights come in FP8 and BF16 formats, under a license that excludes the EU, the UK, and South Korea.
Read original →NVIDIA submitted an updated Vera Rubin NVL72 configuration running DeepSeek V4.1 Flash. At 260 tokens per second per agent, its per-GPU throughput is 5.0× that of GB300; at 80 tokens per second, it is 1.7×. Single-agent speed reaches 1,249 tokens per second, or 1.47× GB300's 848 tokens per second, with gains varying by speed conditions.
Read original →🔍 Analysis: ERNIE-5.5-Flash-Base's 280B total parameters and 14B active parameters describe its scale and computation path. Vera Rubin's 5.0× and 1.7× throughput gains when running DeepSeek V4.1 Flash depend on speed conditions. These figures cannot simply be combined into an estimate of deployment benefits. Enterprise model selection requires checking adaptation work, target generation speed, and license coverage together, rather than judging costs solely by active parameter counts or peak throughput multiples. For ERNIE in particular, the absence of mid-training or post-training means the weights are a starting point for adaptation, not an assistant already optimized for a domain.