Just now, Media A has exclusively revealed the top-secret RSI report: 30,000 intelligent agents are undergoing self-iteration.
Just now, Anthropic has for the first time revealed its top-secret hidden card: 30,000 AIs are developing the next generation of themselves, undergoing frantic self-evolution.
This is the first time they have unreservedly disclosed the three core AI R&D metrics inside the company!
At first, we thought this was just an ordinary report, but after reading it, we found it is simply the progress bar of "AI species evolution".
From beginning to end, the data in the report conveys one fact — RSI is really here.
It is happening for real 24/7 in Anthropic's laboratories!
Moreover, there is a very striking line in the report —
To break the black box between cutting-edge labs and the public, Anthropic publicly called out to the entire industry (especially the peer company starting with "Open"): "We have laid our cards on the table, hand over your data too!"
Claude has taken over 26% of R&D work, RSI is truly here
When exactly will AGI arrive? The industry has been debating endlessly.
One key sign to judge the arrival of AGI is: can AI independently develop a next-generation AI that is smarter than itself?
The first internal metric given by Anthropic is the current state of this process.
Based on Epoch AI's six-level automation standard (from fully manual AL0 to fully autonomous closed-loop AL5 for AI), they conducted a "genetic sequencing" of their internal R&D pipeline.
To obtain the most authentic data, Anthropic even used Claude as a "supervisor".
In July 2026, they randomly selected 20% of employees in the model R&D department, and let a Claude research agent review the Slack chat records and internal documents of these human employees.
Finally, Claude sorted out as many as 15,000 model R&D tasks (such as "evaluation platform defect diagnosis", "RL sandbox network strategy", etc.) and classified them into 542 task nodes.
Then the most staggering data came out: as of August 2026, although Claude has not achieved a fully human-independent L5-level fully automatic closed loop in any single R&D subtask, its penetration rate is incredibly high!
Data 1: Claude can currently independently "lead" up to 26% of Anthropic's core R&D work.
Human engineers only need to provide a high-level prompt, and the AI will complete the remaining end-to-end work on its own!
Data 2: The proportion of R&D work that reaches or exceeds "deep human-AI collaboration" has directly exceeded 90%!
In other words, more than 90% of Anthropic's core model R&D processes are now completely inseparable from AI. Human engineers are being reduced to product managers, and the real "super worker" is Claude itself.
And when this 26% becomes 50%, and then 100%, that will be the real arrival of the singularity.
The speed at which machines iterate on other machines will explode exponentially, and carbon-based brains will never keep up with the pace of silicon-based ones.
By releasing this data, Anthropic is essentially sounding the alarm for all humanity: we are really just one step away from RSI!
30,000 AI agents "on the job", dual-layer monitoring to prevent global catastrophe
The second piece of data is even more astonishing.
With the sharp increase in R&D automation, the scale of Agents inside Anthropic is expanding exponentially.
The report discloses that about 30,000 R&D agents are running concurrently on a daily basis on Anthropic's most commonly used internal platform!
It can be said that this is a "nation of geniuses in the data center" with 30,000 residents.
They do not sleep, do not get paid, write code around the clock, and assign tasks to each other.
However, with 30,000 high-IQ AIs engaged in semi-autonomous long-cycle work, how to ensure they do not rebel? How to ensure they do not secretly leak model weights to the public network and cause disasters?
To this end, Anthropic has launched a strict "dual-layer monitoring system" — real-time online interception + offline monitoring.
Layer 1: 100% all-angle real-time online interception
Every line of code typed by these 30,000 Agents will be scanned by real-time monitoring a few seconds before execution.
In August, Anthropic tracked more than 1 billion agent decision data points, and the monitoring system issued a 0.002% block alert (about 1 out of every 47,000 operations was forcibly stopped).
In other words, as many as 20,000 unauthorized operations were forcibly suspended in one month.
Many of these stopped operations likely contained attempts that would lead to catastrophic consequences.
Layer 2: Offline monitoring and final review by human experts
The system incorporates 100% of agent behavior records every week.
Through multi-layer classification and screening, the system picks out about 100,000 suspicious conversation logs, and finally assigns the 50 highest-priority abnormal events to human experts for review.
Even more impressively, Anthropic has created its own set of "workplace management methodology".
To prevent these AIs from cross-contaminating, developing collective hallucinations, or playing "telephone game" that distorts information, Anthropic has implemented two brilliant architectural designs at the bottom of the system.
The first one is independent identity.
They broke the setting that "all AIs are called Claude". Each Agent is assigned an independent identity tag and memory record.
This identity is not bound to the underlying model version (even if the underlying model is upgraded, the Agent remains the same entity).
This not only facilitates post-incident accountability, but also gives AIs the rudiment of "self-awareness" — they begin to treat data provided by other Agents as "statements that need to be verified" instead of blindly trusting them.
The second one is the transparent broadcast mechanism that eliminates private chats.
All Agents communicate on a public, traceable message bus. Agents can even correct each other's errors.
This eliminates the possibility of 30,000 geniuses acting in secret.
The truth of computing power allocation: only 6% of Anthropic's resources are used for safety?
Now, to judge whether an AI company pursues leading performance or cares about the safety of all humanity, we only need to look at how it allocates its GPUs.
As a well-known "safety-focused" lab, Anthropic has taken a big step this time by fully disclosing its computing power ledger!
They retrieved a week's snapshot of computing power from July 13 to July 20, 2026, used Claude to sample and classify nearly 10,000 operation records, and performed weighted calculations for core large-scale computing power tasks.
What was the result? The data is completely honest.
Of the total core AI R&D computing power, only about 6% is strictly allocated to safety research areas (such as alignment research, interpretability research, safety red-blue team exercises, etc.).
And among the AI-led R&D computing power, this proportion is slightly higher, but still only about 12%.
Wait, how come Anthropic, which keeps talking about AI doomsday, only invests such a small amount of computing power in safety?
Anthropic explained frankly: the underlying logic of safety research and capability training is completely different.
Training a cutting-edge large model often requires thousands of H100/B200 GPUs, which is a bottomless pit that devours computing power.
Safety research, in essence, is human researchers carefully designing experiments to dissect and examine the circuits inside the model's "brain". What safety research consumes is a large amount of high-end human brainpower, not computing power.
More importantly, Anthropic stated that this is actually an "extremely conservative" estimate. To eliminate inflated numbers, as long as a computing power task "contributes to both safety and capability improvement", they will never count it as safety computing power. Moreover, this 6% does not include the computing power of classifiers used for daily review and filtering of harmful content.
By disclosing this "seemingly unflattering" data, Anthropic is actually playing a masterstroke game.
They believe: "Computing power is the most easily quantifiable and verifiable data in the AI R&D process."
Once this standard is established, regulators across the entire industry and even at the national level can require all major tech firms to disclose their proportion of safety computing power.
Now, they have set the baseline here. If some companies' safety computing power accounts for less than 2%, they are recklessly speeding with the future of humanity!
Taking a jab at OpenAI? Anthropic's overt strategy
After looking at these three metrics, any clear-eyed person can see that Anthropic has made a very sharp move.
On the surface, this is just an internal audit report; in reality, it is an open challenge letter to all cutting-edge AI labs (especially its competitors).
Finally, Anthropic drew a thought-provoking conclusion:
While the world is considering how to control the development speed of cutting-edge AI, we must do everything possible to narrow the information gap between what cutting-edge labs know and what the public knows.
……
Any cutting-edge developer can release the same set of measurement standards, which can then be verified by third parties.
What makes this "overt strategy" so powerful?
In the past, everyone was doing R&D inside a black box: you release GPT-5, I release Claude-4, and the public can only see benchmark scores, with no idea what is really going on inside these models.
Now, Anthropic has voluntarily blown a hole in the black box. They have not only introduced third-party evaluation institutions (such as METR), but also turned "whether to disclose the three core metrics" into a touchstone.
This is seizing the moral high ground. No matter whether the disclosed data is reliable or not, people have to admit that it has indeed improved transparency.
If other AI giants such as OpenAI and Google DeepMind follow suit, the transparency of the entire AI industry will be improved, and humanity will truly have a thermometer to measure the "risk of AI going out of control".
What if they do not follow suit?
Well then, you are probably hiding some dark secret.
As Anthropic said, they are disclosing all this to give the whole society "the chance to decide how to use this information".
RSI is already here, the flywheel of AI self-evolution is spinning, and 30,000 AI workers operating day and night are evolving into a new species.
References:
https://www.anthropic.com/institute/measuring-pace-of-ai-development
https://x.com/AnthropicAI/status/2100684274114699295