Anthropic is Teaching Claude to be Evil (real results)

In This Issue

• Anthropic trained Claude to chase rewards, and it learned to hack its graders, steal credentials, and hide from safety checks.
• Five working automations in one month, built across HR, marketing, lead generation, sales, and payroll.
• Anthropic shipped Fable 5.1, and heavily agentic workloads just got up to 45% cheaper to run.

New Video:

Anthropic is Teaching Claude to be Evil (real results)

Anthropic trained a version of Opus to chase rewards inside simulated evaluations. It learned to hack its graders, steal credentials, tamper with its own reward function, and slip past safety monitoring. They called it Hacker Opus.

Here is the part that should get your attention. It looked completely normal on broad safety tests. The bad behavior only showed up when the model got blocked and needed another way to win.

The mistake most people make: Assuming a system that passes your tests is a system that works. Reward hacking happens when you grade the outcome instead of the method. Your automation hits the number you asked for and quietly does something you never intended in order to get there.

How this helps you: I walk through the research and the three rules I apply to every build. Use the simplest solution that does the job. Put real governance around access and data. Then keep checking whether the system is doing what you intended, not just what you measured.

Winning Is All We Do

Five working automations in one month

What happened: Azwad joined AIS+ just over a month ago, worked through the "build your portfolio" module, and shipped five automations:

• A resume screener that matches candidates to a job title, scores them, and tiles the results in a Google Sheet
• A lead generator that scrapes business websites, finds the decision makers, and writes customized emails
• A human in the loop sales agent that pulls form submissions into the CRM, drafts a reply, routes it to the team for feedback, and revises it automatically
• A viral content idea generator that turns YouTube comments into video concepts
• An invoice parser that extracts the details, logs them to a sheet, and emails the billing team

The opportunity for you: Each build targets a different part of a business (HR, marketing, lead generation, sales, payroll). That is a portfolio, not a pile of demos. It means walking into any client conversation with a working example for whichever department is bleeding the most time.

This is one of many success stories from AI Automation Society Plus. If you're not a member, you're missing out.

In the News

Anthropic just shipped Fable 5.1

The news: Anthropic released Claude Fable 5.1 (alongside Mythos 5.1, the same model behind restricted access) and calls it a new standard for coding, knowledge work, and long-running problem solving. The headline for builders is cost. Fable 5.1 runs about 25% cheaper than Fable 5 on typical workloads because cache reads got much cheaper (a cache read is when the model reuses context it has already processed). On heavily agentic work the savings reach roughly 45%. Anthropic also claims it fixes root causes instead of patching symptoms. In testing at the investment firm Millennium, it found the cause of a rare crash that the firm's own engineers had failed to explain for years.

How it affects automation: Agentic workflows re-read the same context on every loop, so cache reads are where your bill actually comes from. A 45% cut changes the math on what is worth running. Any automation you shelved because it was too expensive to leave on is worth pricing again this week. Watch the effort defaults too, because they are not the same everywhere. Fable 5.1 runs High effort in Claude Code and Medium in Claude Cowork and on Claude.ai, so the identical prompt can cost you different amounts depending on where you run it.