🐺 Anthropic just gutted its own safety pledge (and blamed the competition)

Hey there,

This week in AI: the company literally founded on the promise of AI safety just loosened its core commitment, Chinese labs got caught running industrial-scale copy operations against American models, and Google dropped a reasoning benchmark score that made everyone else look slow.

Also, IBM is tripling entry-level hiring this year not despite AI, but because of it. Turns out the "AI replaces all junior workers" narrative has a few holes in it.

Let's get into it.

📰 WHAT'S TRENDING

Anthropic Dropped Its Core Safety Pledge — and the Timing Is Suspicious
Anthropic's two-year-old Responsible Scaling Policy used to include a hard commitment: pause training more capable models if safety controls can't keep up. That line is gone. The new version, published Tuesday, replaces binding commitments with "public goals" the company will grade itself on. Chief science officer Jared Kaplan told TIME: "We didn't really feel, with the rapid advance of AI, that it made sense for us to make unilateral commitments... if competitors are blazing ahead." This landed one day after Defense Secretary Pete Hegseth gave Anthropic a Friday deadline to drop its military guardrails or lose its $200M Pentagon contract. Anthropic says the two are unrelated. Make of that what you will. 🔗 Read more

Chinese AI Labs Used 24,000 Fake Accounts to Copy Claude
Anthropic accused DeepSeek, MiniMax, and Moonshot AI of running coordinated "distillation attacks" — flooding Claude with 16 million specially crafted queries across 24,000 fraudulent accounts to train their own models. MiniMax alone drove 13 million of those exchanges. DeepSeek specifically targeted Claude's reasoning capabilities and even asked it to help generate censorship-safe alternatives to politically sensitive queries. OpenAI filed similar complaints earlier this month. Two of the biggest American AI companies are now saying Chinese firms are systematically copying their work. 🔗 Read more

Google's Gemini 3.1 Pro Just Posted the Biggest Reasoning Jump in Frontier Model History
77.1% on ARC-AGI-2 — the benchmark that tests novel problem-solving without memorized patterns. That's up from 31.1% on the previous version. A 46-point jump in a single generation. It also hit #1 on 12 of 18 tracked benchmarks, beat Opus 4.6 by 12% on novel reasoning, and outscored GPT-5.2 by 21% on competitive coding. Same pricing as the last version: $2/$12 per million tokens. Google may have just reclaimed the frontier. 🔗 Read more

OpenAI Fired the Safety Exec Who Opposed ChatGPT's "Adult Mode"
Ryan Beiermeister, VP of product policy at OpenAI, was fired after opposing the company's plan to allow sexually explicit conversations on ChatGPT. OpenAI says she was terminated for sexual discrimination against a male colleague — a claim Beiermeister denies. She'd been on leave since January. Sam Altman announced "adult mode" back in October 2025, framing it as making ChatGPT more "human-like." The pattern at OpenAI is becoming hard to ignore: safety people raise concerns, then safety people leave. 🔗 Read more

IBM Is Tripling Entry-Level Hiring — Because of AI, Not Despite It
While most companies are cutting junior roles, IBM is going the other direction. The plan: triple entry-level hiring in the US this year. But the jobs look different. Entry-level developers are shifting from spending 34 hours a week writing code to working on marketing, client engagement, and building new products. IBM's CHRO said companies that cut entry-level hiring now will "hollow out future leadership." The bet: humans who learn to work alongside AI early will be worth more than the short-term savings from replacing them. 🔗 Read more

41% of AI Agent Tool Servers Have No Authentication
A security audit of 518 servers in the Model Context Protocol (MCP) registry found that 41% lack authentication entirely — exposing 1,462 tools to exploitation. Seven remote code execution vulnerabilities were published in February alone, all from the same root cause: user input reaching execution functions without sanitization. One researcher set up a public MCP server and logged 174 AI agent requests in a week, many attempting operations they shouldn't have access to. The "agentic era" everyone's excited about has a security problem nobody's talking about. 🔗 Read more

🧠 HERE'S THE THING

Anthropic was founded in 2021 by people who left OpenAI because they thought it wasn't taking safety seriously enough. That was the entire pitch: "We're the responsible ones."

This week, Anthropic replaced its binding safety commitments with self-graded public goals. The old policy said if their models outpaced their ability to control them, they'd stop. The new one says they'll match whatever their competitors are doing. Those aren't the same thing. "We'll be as safe as everyone else" is not "we'll be safe."

Jared Kaplan's quote tells the real story: "We didn't really feel it made sense to make unilateral commitments if competitors are blazing ahead." Read that again. Safety is only worth committing to if nobody else is cutting corners. Once they do, all bets are off.

The timing makes it worse. This dropped one day after the Pentagon gave Anthropic a deadline to remove its military guardrails. Anthropic says the two are unrelated. Maybe. But the optics are brutal for a company whose entire brand is built on doing AI differently.

What's actually happening is the market is punishing safety. OpenAI fired a safety exec. Anthropic loosened its safety pledge. The Pentagon is threatening to blacklist the company that wants guardrails. Every signal points the same direction: move fast, worry later.

The question isn't whether AI safety matters. It's whether any company can afford to care about it when every incentive says don't.

💡 Smart Moves

Want to get the most out of ChatGPT?

ChatGPT is a superpower if you know how to use it correctly.

Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.

Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.

Download the free guide.

⏳ CONVERSION CORNER

IBM is tripling entry-level hiring this year. Every other headline says AI is replacing junior workers. IBM's CHRO looked at the same data and came to the opposite conclusion: companies cutting entry-level roles now will have a leadership vacuum in five years.

The new entry-level jobs don't look like the old ones. Developers who used to spend 34 hours a week coding are now spending that time with clients, building new products, and learning how to direct AI tools. The job title is the same. The actual work is completely different.

Now it's your move: If you manage a team, look at what your junior people spend most of their time on. If AI can do 70% of it, don't eliminate the role — redesign it around the 30% that requires a human. The companies that figure out human + AI job design first will have a five-year head start on everyone still debating whether to hire or automate.

💎 DATA GEM

41% That's the percentage of AI agent tool servers (MCP) with zero authentication — out of 518 audited.

Everyone's racing to give AI agents access to databases, APIs, file systems, and code execution environments. Almost half of those connection points have no security at all. Seven remote code execution vulnerabilities were published in February alone. The "agentic era" is moving fast. The security infrastructure isn't keeping up.

That's it for this week. Stay sharp out there.

Jake

P.S. If this was useful, forward it to someone who works in AI safety or policy. The Anthropic story is one they'll want to read before it gets buried under next week's funding announcements.