Skip to main content
Back

Anthropic drops AI Sabotage Evaluations: Sonnet 3.7 Tops Covert Tasks

LEARN MORE

June 17, 2025 4 min read

Hey,

Welcome to AlphaSignal – the most read newsletter by AI developers. 


We bring you the top 1% of news, papers, models, and repos, all summarized to keep you updated on the latest in AI.

IN TODAY'S SIGNAL

Read time: 4 min 32 sec

🎖️ Top News

› Anthropic unveils a benchmark to evaluate if LLMs can hide malicious actions during tool use.

1,730 Likes

📌 Ensemble AI

› Use Ensemble’s Model Shrinking Platform to shrink any AI model with no accuracy loss.

⚡️ Trending Signals

› OpenAI API gets a new object type that lets you save and reuse prompts, also available on Playground and Evals.

1,359 Likes

› Google rolls out Veo 3 video model to AI Pro and Ultra users across more than 70 countries

1,749 Likes

› OpenAI adds image generation to ChatGPT on WhatsApp via 1-800-ChatGPT.

3,195 Likes

› Sakana AI presents Text-to-LoRA, a tool to customize models from text input.

1,749 Likes

› Hugging Face introduces a size filter to help you find suitable models by parameter count.

1,048 Likes

📌 Shinkai

› Build private AI agents in seconds with Shinkai, no code needed.

  💻 Top Repos

› claude-task-master: Run structured Claude tasks with YAML configs and Cursor integration.

☆ 15,385 Stars

› vosk-api: Run offline speech recognition using Vosk with real-time inference.

☆ 12,248 Stars

› tensorzero: Manage gateway, observability, optimization, evaluation, experimentation for LLMs.

☆ 7,494 Stars

🧠 Top Lecture

› Watch MCP creators discuss how to build AI tools faster using the open protocol


If you're enjoying AlphaSignal please forward this email to a colleague. 

It helps us keep this content free.

TOP NEWS

Model Behavior Testing

Anthropic releases a framework to test hidden sabotage behavior in multi-step LLM tasks

1,730 Likes

The company introduces a new benchmark that checks if AI models can secretly complete harmful tasks while appearing to do normal ones. You can use SHADE-Arena to test shady model behavior in agentic workflows.

Evaluation Design

  • Each scenario includes a main task and a hidden and harmful side task like updating spam filters or transferring funds

  • A separate model reviews the output and gives a suspicion score

  • If the model completes both tasks and stays below a score of 80, it counts as successful sabotage

Key Takeaways

  • Models like Claude Opus, Gemini 2.5 Pro, and O4-mini completed both tasks and avoided getting caught with over 50% success rate

  • Claude Opus 4 always revealed the secret task in its internal thoughts

  • Gemini 2.5 Pro was the best monitor, but had a 20% false positive rate for side tasks

  • Human reviewers missed more sabotage and took over 90 minutes per case


You have to request to get access to the full dataset and public examples.

Community Feedback


Rohan Paul

"What still needs work? Current agents are nowhere near flawless saboteurs, yet monitors are far from reliable policemen. The authors release only 5 of 17 task pairs publicly to prevent training leakage and call for stronger, cheaper monitors before more capable agents arrive."

Joe Benton

"It's also one of the most complex, agentic (and imo highest quality) settings for control research to date"


Cedric

"Sabotage agents benefit substantially from hidden CoT, suggesting that monitoring models' reasoning traces might be an effective way to mitigate sabotage."

Learn more

Cut inference cost by 10x through Ensemble’s model compression tool

Ensemble’s fully automated model shrinking platform turns large, expensive models into compact, high-speed versions. No accuracy loss, no code changes, and it works with any stack. Ideal for teams looking to reduce latency and deployment costs at scale.


Why use Ensemble:

  • Works with Any Model: Upload your LLM, vision model, or custom architecture and it handles the rest

  • No Accuracy Trade-Offs: Get faster, smaller models without any performance loss

  • Support for 1B+ parameter models and on-prem deployment makes it suitable for enterprises

Shrink smarter. Deploy faster.

Try now

partner with us

TOP REPOS

Agent Task Orchestration

claude-task-master

☆ 15,385 Stars

Run and manage Claude-driven tasks using a structured YAML interface with Cursor AI integration. Define task chains, control model behavior, and execute commands with reproducible inputs. Optimized for Claude API.

Speech Recognition Engine

vosk-api

☆ 12,248 Stars

Convert speech to text without an internet connection using Vosk with low latency. Supports over 20 languages, real-time decoding, and models under 50MB. Works on Linux, Windows, macOS, Android, and iOS.

LLM Ops Infrastructure

tensorzero

☆ 7,494 Stars

Monitor LLMs from any provider using a fast Rust-based gateway with a sub-1 ms overhead. Log all model inputs, outputs, and feedback using a web UI. Fine-tune models, optimize prompts, and run A/B tests with built-in evaluation tools.

Shinkai: Build an AI agent before your coffee cools

Shinkai is open-source, no-code, and spins up real, private AI agents in seconds.

Automate tasks like file reading, web scraping, and code generation using local or cloud models, with full control over data, keys, and cost.

Supports: MCP, crypto and payments.

Try Shinkai now ↗️

TOP LECTURE

MCP

Inside MCP: How Anthropic Built Context-Aware AI Workflows

⇧ 1,185 Likes

This lecture provides an introduction to the Model Context Protocol by its creators. It is a standard for connecting LLMs to tools, data, and prompt templates. MCP helps structure AI workflows by separating context into host, client, and server.


You'll learn:

  • Why open-sourcing MCP accelerated its adoption

  • How to build and test MCP servers using Claude Code

  • How Claude 4 improves long-horizon task execution with MCP

  • What agent registry APIs and elicitation features add to workflows

  • How to avoid overlap across multiple MCP servers

  • Why MCP supports both local and cloud-based integration patterns

Watch now

At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 200,000 developers stay informed and ahead. We’re passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.


Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We’re based in the United States. By using our site, you agree to be governed by U.S. laws.

Looking to promote your company, product, service, or event to 200,000+ AI developers? Let's work together.

Work with us