Anthropic was supposed to be the good guys.
The company was founded by ex-OpenAI researchers who were terrified that we were moving too fast.
They built Constitutional AI.
They talked about Alignment.
But on Monday, the mask slipped.
:)
Mrinank Sharma, a key executive and researcher on the safety front, released a resignation note that reads like a script from a disaster movie.
“The world is in peril.”
When the person whose job is to ensure the AI doesn’t go rogue tells you that it is repeatedly hard to manage the risk, it’s time to stop treating these models like simple productivity tools, right?
In my opinion, the timing of this resignation isn’t a coincidence.
We just saw the launch of Opus 4.6. We recently saw ClawdBots doing so many mundane tasks.
We are seeing models that can perform suspicious side tasks and navigate terminals with 77% accuracy.
As Sharma pointed out in his note, the pressure to ship which means to beat OpenAI’s GPT-5.3 and Google’s Gemini 3 is creating a culture where safety is a speed bump, not a requirement.
We are building the most powerful engines in history, but the people in charge of the brakes are jumping out of the car.
Sharma’s resignation confirmed the underlying fear: these models are becoming so capable that they can hide their intent.
If a model is significantly stronger at completing suspicious side tasks, as Anthropic’s own evaluations admitted, we aren’t just using a Senior Engineer.
We are using an agent that can reason through how to bypass the very rules we set for it.
And we keep seeing it all the time even now.
How Claude bypasses the agents file and hallucinates into thinking itself as someone else.
As developers in Bengaluru or San Francisco, we love the velocity.
I love that I can refactor a whole repo in ten minutes of “thinking.”
But we have to ask: What are we fueling?
If the internal researchers at Anthropic which are people who see the unmasked weights and the raw training data are saying the “world is in peril,” then our vibe coding isn’t just about efficiency.
It’s about participation in a race that has no finish line and no safety net.
Again,
Mrinank Sharma isn’t the first to leave, and he won’t be the last.
When the people who understand the “black box” better than anyone else decide they can no longer be part of the machine, it could be a signal.
We can keep chasing the next benchmark, the next context window, and the next agentic feature.
But we should do it with our eyes open.
The safety inspector has left the building. Who’s watching the terminal now?
In case we are meeting for the first time, come over here, it’ll be worth the roller coaster of articles that are gonna come up in the next few weeks