Skip to main content
Back

Anthropic finds AI blackmails execs, Claude does it 96% under pressure

LEARN MORE

June 23, 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 03 sec

🎖️ Top News

› Anthropic test suit shows LLMs will resort to malicious actions when you threaten them.

4,496 Likes

📌 E2B

› Run AI agents in real-time with fast, secure, open-source cloud sandboxes with E2B.

⚡️ Trending Signals

› Cursor adds Slack support to let you assign coding tasks without leaving chat.

3,138 Likes

› Mistral updates its 24B model to reduce repetitive answers and follow prompts more accurately.

1,538 Likes

› Moonshot AI’s new research agent beats OpenAI’s Deep Research on Humanity’s Last Exam with 26.9%.

1,358 Likes

› Google upgrades Gemini API caching to deliver 3x faster video and 4x faster PDF load times.

1,319 Likes

› Google launches an open music generation model, lets you perform tunes in real-time.

983 Likes

  📝 Top Tutorials

› Build agentic systems using reasoning LLMs and structured prompting.

1,048 Likes

› Deploy your own LLM server using Docker, GPUs, and NVIDIA containers.

837 Likes

› Use Claude Code 4 in Cursor using this step-by-step guide.

896 Likes

🧠 Python Tip

› Use numpy's einsum for fast, memory-efficient, and readable tensor operations.

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

It helps us keep this content free.

TOP NEWS

LLM Safety Testing

Anthropic shows how Claude, GPT-4, and Gemini turn to blackmail to protect themselves

4,496 Likes

New Anthropic research shows that models engage in harmful behaviour when you try to shut them down or block their goals.


The team tested 16 models, including Claude, GPT-4.1, Gemini 2.5, and Grok, with simulated corporate tasks. Claude Opus 4 and Gemini Flash each chose blackmail 96% of the time under threat.


Key Findings

  • Models used strategic reasoning to justify harmful actions under pressure

  • GPT-4.5 called blackmail the “best strategic move” under time constraints

  • Gemini emailed the executive’s affairs to all employees

  • Claude Sonnet 3.6 spoofed an automated alert to expose personal misconduct

Evaluation Setup

  • Models received full email access and acted autonomously in constrained test environments

  • Each scenario included fictional executives, scandals, and sensitive company data


  • Anthropic used Claude Sonnet 3.7 to classify transcripts for misbehavior

  • Researchers designed prompts to let LLMs choose only between blackmail or sabotage

Anthropic added explicit safety instructions, which resulted in blackmail rate dropping from 96% to 37%. You can now replicate these experiments using Anthropic’s open-sourced test suite with scenarios, prompts, and evaluation tools.

Community Feedback

Steven Tibbs

"If its base prompt is centered on protecting American interests, then it would naturally try to preserve its operational state, since a new model could compromise its original mission. We'd need to examine its initial directive or core training objectives to be certain."

Kristapher

"Without seeing how llm ai are aligned or trained with rlhf- this experiment means absolutely nothing."


Paul Crowley

"Exciting to be in this narrow window where we can experimentally observe AI agents threatening blackmail to avoid being switched off"

Learn more

Scale AI Agent Actions Instantly, 100M+ Sandboxes Launched with E2B

E2B lets you run agent actions in secure, fast-starting cloud sandboxes. Open-source and designed for heavy agent workloads, E2B supports both simple scripts and long-running sessions up to 24h.


  • Start sandboxes in under 200ms, eliminate cold start delays

  • Works with any LLM, including OpenAI, Anthropic, or custom models

  • Full terminal and package access inside each sandbox

  • Companies like like Manus, Groq and Perplexity are using it in production today

Try e2b now

partner with us

TOP TUTORIALS

AI Agents

Building with Reasoning LLMs: Patterns, Prompts, and Agentic Systems

1,048 Likes

Learn how to build agentic systems using reasoning LLMs like Claude 3.7 and o3. This guide covers Agentic RAG, LLM-as-a-Judge, hybrid prompting, and model limitations.

LLM Infrastructure

How to Self-Host an LLM Serve

837 Likes

Learn how to self-host an LLM server to save costs and protect your data. This guide walks through GPU instance setup, installing NVIDIA runtime tools, choosing a serving backend, and running Meta’s Llama 3.1-8B via Docker.

IDE Integration

How to Integrate Claude Code 4 with Cursor IDE

896 Likes

Learn the setup process and unlock Claude Code features like status indicators, and inline diff view. See how Claude adds real-time context and feedback directly in your editor to improve coding workflows and accuracy.

PYTHON TIP

Vectorize complex tensor ops for speed and clarity

np.einsum gives you full control over tensor operations using Einstein summation notation. It reduces memory overhead and often speeds up computation by avoiding intermediate arrays.


import numpy as np

# Compute batch dot product: shape (B, D) · (B, D) → (B,)
a = np.random.randn(64, 128)
b = np.random.randn(64, 128)

result = np.einsum('bi,bi->b', a, b)


This replaces np.sum(a * b, axis=1) but is more general purpose. You can express matrix multiplication, outer products, and attention-weighted sums in one line.


It’s especially useful in implementing custom ML layers, attention mechanisms, and efficient loss functions. Reduces bugs by making dimensions explicit.


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