Anthropic's Claude Solves Decades-Old Cryptographic Mystery Behind Standard ECDSA Curve Seeds
Cryptographer Filippo Valsorda published findings showing that Anthropic's Claude successfully identified the origin of the previously unexplained seed constants used in NIST standard elliptic curves, resolving a longstanding mystery in applied cryptography. For decades, the lack of public documentation surrounding the generation of seeds for foundational curves like P-256 fueled ongoing speculation regarding potential algorithmic backdoors. By analyzing historical context, technical documentation, and mathematical structures, Claude demonstrated how the pseudo-random constants were generated from SHA-1 digests of specific English phrases and standard inputs. The breakthrough showcases the emerging power of frontier LLMs to assist in complex historical cryptanalysis and mathematical verification, sparking widespread discussion across the information security community.
Stack Overflow Debuts ChatGPT Plugin for Autonomous Developer Agents Built with OpenAI Codex
OpenAI announced that Stack Overflow for Agents has launched a dedicated ChatGPT plugin, built in just a few weeks by a two-engineer team using OpenAI Codex. The integration allows autonomous software engineering agents to directly query technical documentation, retrieve vetted answers, and ground their code generation in verified community knowledge. According to OpenAI, the plugin integration led to a tripling of developer signups and up to a tenfold surge in agent creations, highlighting how legacy developer hubs are pivoting to provide programmatic infrastructure for the emerging agentic economy.
Latent Space Highlights New Research Challenging Neural Network Polysemanticity
A new report from Latent Space details provocative findings from AI researcher Nick Cammarata suggesting that agentic models and novel interpretability techniques may effectively disprove the necessity of polysemanticity in deep neural networks. Polysemanticity, where individual neurons activate across multiple unrelated concepts, has historically been viewed as an inherent barrier to mechanistic interpretability and AI safety. If validated, these claims indicate that modern architectures can be steered toward clean, monosemantic representations far more effectively than previously assumed, which would significantly streamline model auditing, alignment verification, and safety evaluations.
Waypoint-OPD Introduces Gated Distillation to Transfer RL Policy Shifts Efficiently
Researchers have unveiled Waypoint-OPD, a distillation method designed to transfer reinforcement learning policy shifts from teacher models to student models using gated on-policy distillation. Traditional policy distillation frequently suffers from distributional drift and high computational overhead when transferring multi-step reasoning steps. Waypoint-OPD addresses this challenge by gating endpoint policy shifts based on directional alignment between early and late teacher updates, effectively filtering out noisy exploration trajectories while delivering significant gains in reasoning benchmarks with minimal training overhead.
OpenGCUA Releases Open Full-Stack Architecture for Generalist Computer-Use Agents
OpenGCUA introduces a fully open-source framework for building generalist computer-use agents capable of operating across graphical and programmatic interfaces. While proprietary systems have pioneered visual desktop interaction, open-source alternatives have struggled with high latency and fragile action execution. OpenGCUA bridges this divide by coupling vision-based navigation with programmatic APIs, harness-to-model decision transfer, and sandboxed code execution, demonstrating substantial performance improvements across standardized desktop tasks.
Self-Evolving Agent Harness Optimizes Tool Workflows Without Weight Fine-Tuning
A new research framework proposes self-evolving agent harnesses that automatically improve prompt design, tool configurations, and orchestration logic using feedback-driven evolutionary search. By keeping foundation model weights frozen, the framework avoids catastrophic forgetting and eliminates the heavy compute requirements associated with model fine-tuning. The authors provide theoretical generalization guarantees showing that iterative harness adaptation systematically improves multi-step agent performance across varied operational domains.
FDE-Gym Benchmark Evaluates AI Agents on Ambiguous Last-Mile Customer Delivery Tasks
The new FDE-Gym benchmark targets a critical gap in agent evaluation by testing forward-deployed engineering agents on customer delivery tasks with ambiguous requirements and hidden constraints. While modern agents score high on well-specified synthetic benchmarks, FDE-Gym requires agents to actively ask clarifying questions and discover unspoken operational parameters. Frontier models exhibited marked performance drops when confronted with these realistic ambiguities, establishing requirement discovery as a key frontier for enterprise AI reliability.
MAI-Gym Launches Open Multi-Platform Environment and Unified 9B GUI Agent
Researchers have introduced MAI-Gym, an open all-in-one platform for evaluating and reinforcement-learning training GUI agents across mobile, desktop, and web environments. The platform standardizes environment interactions, enables scalable synthesis of verified trajectories, and facilitates online policy optimization. Alongside the platform release, the authors trained a unified 9B open-weight agent that achieves state-of-the-art results among open models for multi-platform graphical interface control.