Tootfinder

OpenAI says it plans to let third-party groups conduct technical safety evaluations of its AI models during the training, evaluation, and deployment phases (Rachel Metz/Bloomberg)
https://www.bloomberg.com/news/articles/2026…

Before everybody get's too excited about available Chinese open weights models, the infrastructure costs for running these models are immense, it's hard to find specific and reliable data but think about half a million dollars of infrastructure costs for 5-15 users according to this linkedin post.

Kimi K3 just dropped. 2.8 trillion parameters, 1M-token context window, open weights. Capabilities comparable to Fable 5. 🔥 And every enterprise buyer is doing the same math right now. Here is… | Miklos Toth
Kimi K3 just dropped. 2.8 trillion parameters, 1M-token context window, open weights. Capabilities comparable to Fable 5. 🔥 And every enterprise buyer is doing the same math right now. Here is what running it yourself actually costs: - 16x NVIDIA H200 GPUs: ~$500,000. That's GPUs only. No chassis, no networking, no power infrastructure. - Model weights on disk: 1 to 1.7TB depending on quantization. - What's left for actual context: somewhere between 550 and 1,250GB of KV cache. Translat…

Leaky Language Models: Stealing Architecture and Inference Optimizations via Per-Token Timing
Sadegh Majidi, Niloofar Mireshghallah, Kazem Taram
https://arxiv.org/abs/2607.20723 https://arxiv.org/pdf/2607.20723 https://arxiv.org/html/2607.20723
arXiv:2607.20723v1 Announce Type: new
Abstract: This work presents LeakyLMs, a set of attacks that leak proprietary model, architecture, and deployment information from production language models. LeakyLMs is the first to demonstrate that key model and deployment details can be inferred using only token generation timing, even when interacting through remote APIs. LeakyLMs introduces two core attacks. The first attack targets inference optimizations and deployment strategies. For example, our attack detects whether a provider uses speculative decoding, a widely deployed inference-time optimization, and further identifies the context length of the draft model used in the pipeline. Our measurements show that Google Gemini Flash 2.5 uses speculative decoding with a draft context window of approximately 128K tokens. The second attack recovers key architectural properties, including the number of transformer layers, hidden dimension size, and number of attention heads. To achieve this, LeakyLMs builds a detailed and accurate model of token-generation timing on modern NVIDIA GPUs, characterizing how latency scales with model configuration and hardware parameters. The attack then performs a search over the architecture space using this timing model. In experiments with Llama models, the near-correct architectural configuration appears in the top-10 guesses more than 90% of the time.
toXiv_bot_toot

Psychedelics align brain activity with context https://www.nature.com/articles/s41586-026-10910-z Under psilocybin, "the separation between internal models and sensory context on which predictive processing depends" dissolves.
Brain-wide reconfiguration of burst firin…

Psychedelics align brain activity with context - Nature
The psychedelic psilocybin reorganizes brain activity into structured, context-aligned patterns that integrate internal and external processing, providing a neural basis for the felt continuity between self and world, elucidating how the psychedelic state translates into psychological change.

Interesting study commissioned by European Commission's DG ENER on smart and bidirectional charging of EVs.
Smart charging is already mature tech and ready for use. It is important to use time-varying energy and network prices, and for TSOs and DSOs to remove barriers to small-scale flexibility.

Study on the assessment of use cases and business models for smart and bidirectional recharging in the context of energy system integration - Publications Office of the EU
demand patterns across the European Union. As electrification progresses across private, commercial, and fleet segments, charging behaviour has become a material factor for peak demand, grid utilization, and system balancing. Within this context, smart charging (SC) and bidirectional charging (BDC) have growing relevance as demand-side flexibility and energy-system integration instruments. This study first assesses the system relevance, maturity, and deployment conditions of SC and BDC across t…

Replaced article(s) found for cs.CR. https://arxiv.org/list/cs.CR/new
[1/2]:
- Facade: High-Precision Insider Threat Detection Using Deep Contextual Anomaly Detection
Alex Kantchelian, et al.
https://arxiv.org/abs/2412.06700 https://mastoxiv.page/@arXiv_csCR_bot/113627226016881063
- Enhancing Membership Inference Attacks on Diffusion Models from a Frequency-Domain Perspective
Puwei Lian, Yujun Cai, Songze Li, Bingkun Bao
https://arxiv.org/abs/2505.20955 https://mastoxiv.page/@arXiv_csCR_bot/114584229384238830
- Cryptographic Choreographies
Sebastian M\"odersheim, Simon Lund, Alessandro Bruni, Marco Carbone, Rosario Giustolisi
https://arxiv.org/abs/2602.12967 https://mastoxiv.page/@arXiv_csCR_bot/116079470877038972
- TALUS: FIPS-204-Exact Threshold ML-DSA via Boundary Clearance
Leo Kao, Raymond Chang
https://arxiv.org/abs/2603.22109 https://mastoxiv.page/@arXiv_csCR_bot/116283533730322814
- SafeHarbor: Defining Precise Decision Boundaries via Hierarchical Memory-Augmented Guardrail for ...
Liu, Ying, Zhang, Zou, Zhang, Yang, Zhang, Peng
https://arxiv.org/abs/2605.05704 https://mastoxiv.page/@arXiv_csCR_bot/116537968700983890
- AI Security Policy Should Assess Systems, Not Only Models
Michael A. Riegler, Inga Str\"umke
https://arxiv.org/abs/2605.09504 https://mastoxiv.page/@arXiv_csCR_bot/116560756584469721
- Efficient and Privacy Aware Edge Cloud Collaborative Inference for Large Language Models
Cheng Li, Jiexiong Liu, Yixuan Chen, Yi Li
https://arxiv.org/abs/2607.13093 https://mastoxiv.page/@arXiv_csCR_bot/116928605403267286
- From Neural Intent to Cryptographic Authorization: Securing AI-Driven Enterprise Workflows
Jiasi Weng, Jian Weng, Minrong Chen, Ming Li, Jia-Nan Liu, Zhi Li, Yue Zhang
https://arxiv.org/abs/2607.15596 https://mastoxiv.page/@arXiv_csCR_bot/116951166143038362
- Towards an Automated Test of LLM Security Knowledge
Shufan Chai, Liangliang Sun, Jessica Staddon
https://arxiv.org/abs/2607.18496 https://mastoxiv.page/@arXiv_csCR_bot/116962579240952448
- DynaMark: A Reinforcement Learning Framework for Dynamic Watermarking in Industrial Machine Tool ...
Navid Aftabi, Abhishek Hanchate, Satish Bukkapatnam, Dan Li
https://arxiv.org/abs/2508.21797 https://mastoxiv.page/@arXiv_eessSY_bot/115128274334268839
- CLOAK: Contrastive Guidance for Latent Diffusion-Based Data Obfuscation
Xin Yang, Omid Ardakanian
https://arxiv.org/abs/2512.12086 https://mastoxiv.page/@arXiv_csLG_bot/115729063749487767
toXiv_bot_toot

SpaceXAI releases Grok 4.7, which it says is better at verifying its own work and managing longer context, available for $2/1M input and $6/1M output tokens (xAI)
https://x.ai/news/grok-4-7

Introducing Grok 4.7
SpaceXAI's most powerful model for coding and knowledge work. Twice as fast, at half the price of comparable models.

PhantomSeal: Proactive Deepfakes Defense with Identity/Context Protection and Forensic Tracing
Liangqin Ren, Zeyan Liu, Ye Wang, Yuxin Chen, Fengjun Li, Bo Luo
https://arxiv.org/abs/2607.20564 https://arxiv.org/pdf/2607.20564 https://arxiv.org/html/2607.20564
arXiv:2607.20564v1 Announce Type: new
Abstract: Deepfakes, especially face-swapping attacks, pose significant challenges to authenticity, security, and ethics across science, engineering, and society. While most existing detection/tracing approaches operate post hoc, proactive defenses that aim to intervene before deepfake generation remain limited in terms of real-world effectiveness. In this paper, we present PhantomSeal, the first proactive defense to simultaneously protect both the identity and the context of users' images from being used in face-swapping attacks, while supporting forensic tracing. We present a novel cloaking technique that embeds a selected identity as a stealthy identifier. This mechanism steers the deepfake generation process toward producing content that resembles the chosen cloak identity, thereby preventing successful face-swapping while enabling effective feature-based forensic analysis. The effectiveness and robustness of PhantomSeal is demonstrated in extensive experiments across different face-swapping architectures and models. For example, it reduces the attack success rate of SimSwap, an advanced deepfake model, to 0.30%, and correctly identifies 97.97% of manipulated content. Codes can be found at https://github.com/LiangqinRen/PhantomSeal
toXiv_bot_toot

Over 30 crypto companies, including Coinbase and Block, say frontier AI safety guardrails hinder legitimate security work while attackers use stronger tools (Shaurya Malwa/CoinDesk)
https://www.coindesk.com/tech/2026/08/13/bitcoin-firm…

Bitcoin firms ask AI labs for same tools attackers already have
Coinbase, Block, BitGo and dozens of others signed a letter arguing that safety guardrails on frontier models are blocking legitimate security work while adversaries face no such limits.

Anthropic is partnering with Accenture to embed evaluators within Anthropic, including red teaming models and conducting alignment assessments (Anthropic)
https://www.anthropic.com/news/accenture-embedded-evaluation

Partnering with Accenture on embedded evaluation
We’re partnering with Accenture on independent evaluation of frontier AI—part of our recent commitment to embed evaluators at Anthropic. Both we and Accenture expect to invest at least $1 billion to build capacity in this area over the next five years.