AI FRONTIER
Signals worth understanding before they become obvious.
Introducing AlphaProof Code: a breakthrough in verifying complex software at depth
Google DeepMind researchers used AlphaProof Code, an agentic framework, to formally verify key security properties of the actual code underlying the Signal protocol, with agents generating over six million lines of Lean.
- AlphaProof Code is a multi-LLM agentic framework in which human experts work with AI to decide what to prove and a group of AI agents does the proofs.
- Agents generated over six million lines of Lean to prove security properties of Signal's cryptographic protocol implementation.
- Human review covered just one percent of the codebase, with Lean automatically checking the rest.
- The verification process took four weeks, a task that historically takes expert teams years.
For AI builders, this shows that multi-agent proving plus Lean's automatic checking can push formal verification from design-level analysis down to compiled production code while concentrating human review on a tiny fraction of the codebase...
Introducing the Multimodal Knowledge Agent benchmark
Google introduces the Multimodal Knowledge Agent benchmark to measure model performance on specific tasks.
- The benchmark is named the Multimodal Knowledge Agent benchmark.
- The source is blog.google.
- The paper abstract states it proposes or uses a new benchmark evaluation to measure model performance on specific tasks.
AI builders can use this benchmark to evaluate multimodal knowledge agents on specific tasks.
OpenRouter adds Baidu ERNIE-Light-DT-1.0
ERNIE-Light-DT-1.0 is based on open-source GLM models and incorporates research findings from Baidu Search's Lightning large model, using Baidu Sparse Attention (BSA) plus feedback-driven history compression to cut KV cache storage and compute cost.
- Context window is 1,024,000 tokens with up to 131,072 completion tokens.
- Pricing is $1.13 per 1M input tokens and $3.94 per 1M output tokens, with cache reads at $0.113 per 1M tokens.
- Hosted by a single provider, Baidu Qianfan, with 83 tok/s throughput, 1.05 s latency, and 100.00% uptime over 3 days.
- Supports function calling via tools and tool_choice, and structured outputs via a JSON schema in response_format.
For builders pushing million-token contexts, BSA and feedback-driven history compression are the levers to test against real long-document and long-session workloads where KV cache cost dominates.
OpenRouter new model: Thinking Machines Inkling 2 Small (free)
Inkling 2 Small is a multimodal mixture-of-experts reasoning model from Thinking Machines Lab with 17B active parameters out of 594B total, and the first model in the second Inkling generation. It was trained with reinforcement learning on long-horizon coding tasks across many agent harnesses, accepts text and image in...
- 17B active parameters out of 594B total.
- 1,048,576-token context window with up to 131,072 completion tokens.
- Accepts text and image input and returns text; no audio support.
- The free endpoint is only available for use with agentic harnesses, and prompts and outputs are logged and used to improve Thinking Machines Lab's models.
Builders get a zero-cost way to test a 1M-context agentic coding model, but the free endpoint's logging terms mean confidential or personal data should not be sent to it.
RL-ATTACK: A Generalized Black-Box Adversarial Attack on Language Models via Reinforcement Learning
The paper proposes a trainable black-box reinforcement learning framework that learns from the target model to generate tailored adversarial transformations instead of guessing transformations from external knowledge. It reports lower post-attack accuracy, greater query efficiency and better semantic preservation on na...
- On natural language inference, the paper reports driving post-attack accuracy to 2% within 20 queries.
- Against generative language models, the trained attacker consistently achieves lower post-attack accuracy than the untrained attacker.
- Compared with a strong optimization-based baseline, the paper reports winning in 4 out of 6 settings.
- The paper experiments with multiple reward designs and training settings and finds no design or setting consistently dominant.
For AI builders, this indicates black-box attackers can be trained from the target model's own signals to expose vulnerabilities with fewer queries for safety evaluation.
Obsidian Security launches runtime policy engine to control AI agent actions
Obsidian Security launched a runtime policy engine to control AI agent actions. The launch was reported by Unite.AI.
- Obsidian Security launched a runtime policy engine to control AI agent actions.
- The launch was reported by Unite.AI.
For builders of AI agents, runtime policy enforcement means agent actions can be constrained at execution time rather than relying only on pre-deployment static configuration.