Kairos Files

What leaked

Four ideas appear repeatedly across the leaked material.

01

Always-on runtime

KAIROS is presented as a system that can remain active across multi-step tasks instead of responding one turn at a time.

02

Persistent memory

The leak describes checkpointed state that can be saved, resumed later, and carried across sessions instead of living inside one context window.

03

Coordinator logic

KAIROS is described as a coordination layer that assigns specialized work to sub-agents and then combines their output.

04

Containment layer

The public response was a patch after a packaging error exposed internal logic that was not intended for release.

Timeline

The leak is easier to understand once the events are put back in sequence.

March 31, 2026 04:20 ET

Exposed source maps are discovered in a public package

Researcher Chaofan Shou identifies publicly accessible source map files, shifting the story from rumor to a documented packaging mistake with inspectable artifacts.

Source: Kuber Studio Type: discovery

March 31, 2026 05:30 ET

The leaked package is reconstructed into a readable codebase

Developers use `npm pack` and related tooling to extract roughly 1,900 files, turning scattered fragments into a readable codebase that can be inspected directly.

Source: smithstephen.com Type: extraction

March 31, 2026 08:00 ET

Mirrors spread and KAIROS becomes a public story

Reconstructed files are mirrored across GitHub and GitLab, and attention shifts from a single discovery to broad circulation of the leaked architecture.

Source: GitHub Type: propagation

March 31, 2026 12:00 ET

Public analysis shifts from the leak to the system design

By midday, discussion focuses less on the exposure itself and more on what the code implies: recursive planning loops, checkpointed state, and delegated sub-agents coordinated by a commander layer.

Source: Medium Type: analysis

April 1, 2026 13:00 ET

Patch confirmation closes the first phase of the incident

Anthropic says it patched the exposure. People are still analyzing the leaked code, but the original packaging mistake appears to have been fixed.

Source: article update Type: response

Evidence blocks

The supporting material becomes clearer when it is grouped into exhibits.

Memory residues Canonical

Persistent state is described as a direct answer to context-window limits.
Checkpointed work can pause, remain dormant, and resume later.
The system is framed as handling tasks that outlast a single live session.

Backed by the article’s checkpointing and persistence framing.

Daemon processes System layer

Recursive planning breaks larger goals into staged task trees.
Multi-agent hierarchy delegates specialized work to sub-agents.
Aggregation recombines sub-agent output into a single response.

Derived from the article’s planner and commander-system framing.

Redacted files Containment layer

The leak is described as the result of a packaging or source-map configuration mistake.
The exposed material was then reconstructed, mirrored, and analyzed in public.
According to the article, Anthropic’s confirmed response was to patch it.

Pulled from the article’s incident and response summary.

Rumor layer

Reddit made the leak broader, stranger, and less dependable.

The Reddit thread expands the leak into a wider mythology, but the article grounds the core story in three clearer mechanisms: recursive planning, persistent checkpoints, and multi-agent delegation.

That makes it easier to separate mood from evidence. The rumor layer still contributes atmosphere, but the article gives the page a firmer technical center.

Final state

The leak ends at the interface, where proactive behavior becomes visible to the user.

Recovered transcript live buffer

system transcript reconstructed from the exposed interface layer.