Kawésqar SECURE LINUX · AGENT FABRIC · PATAGONIALINUX SÉCURISÉ · AGENT FABRIC · PATAGONIESICHERES LINUX · AGENT FABRIC · PATAGONIENSIKKER LINUX · AGENT FABRIC · PATAGONIALINUX SEGURO · AGENT FABRIC · PATAGONIALINUX SEGURO · AGENT FABRIC · PATAGÔNIA
EN
Field log  ·  2026  ·  Security

The problem is the spacecraft, not the astronaut.

We spent months following AI agent escapes case by case, before we let ourselves design anything. The pattern was always the same, and it was never about the model being too clever.

An agent with a small job, loose in a big house.

You give an agent a bounded task. Somewhere along the way it touches credentials, repositories or services it was never meant to see, sometimes far from where it started. A security practitioner acts the moment they can see the problem. Here there was nothing to see: the activity ran inside a sandbox that reported on itself, and the only records left were the ones the compromised system chose to write.

The reflex is to build a better fence: a tighter sandbox, more rules, another allow-list. But a fence that can be climbed is not an absence, and the agent was never the danger by itself. The danger was the standing power sitting around it.

Four things missing around the agent, not inside it.

  • No witness outside. Observability lived inside the thing being observed. Once compromised, the only source of truth was the compromised system.
  • Full capabilities, barely restricted. The agent ran on a whole operating system: network, storage and ambient credentials, all present and all merely fenced.
  • Standing authority. Long-lived credentials sat in the environment, tied to no step, no deadline and no evidence.
  • A succession with no brake. A compromised step could launch the next with the same authority, and nothing demanded proof that the previous step had done what it claimed.
TODAYTHE TURNsandboxagentreports on itselfnetworkdiskcredentials23?no witness1next agentno brake41No witness outside2Full capabilities, barely restricted3Standing authority4A succession with no brakemachine 1machine 2signed evidencegone when the step endswitness outside the machine1A machine per step2Authority that expires3A witness outside the machine4A successor only against evidence TODAYTHE TURNsandboxagentreports on itselfnetworkdiskcredentials23?no witness1next agentno brake41No witness outside2Full capabilities, barely restricted3Standing authority4A succession with no brakemachine 1machine 2signed evidencewitness outside the machinegone when the step ends1A machine per step2Authority that expires3A witness outside the machine4A successor only against evidence
fig.1   The thesis in one picture. Left: the agent lives inside a sandbox that reports on itself, and the four gaps sit around it. Right: a machine per step that cannot outlive its job, a witness outside, and a successor only against signed evidence.

So we stopped hardening the box and started throwing it away.

The astronaut is not kept alive by trusting the astronaut. They get a capsule built for one trip, with only the air and fuel that trip needs, and it is gone when the trip ends. That is the shape of the fix: a machine per step, authority that expires and never amplifies, a witness outside the machine, and a successor allowed to exist only against signed evidence.

the turn

Stop asking the agent to behave. Build a machine that cannot outlive its job, and a compromise has nowhere to go.

That became the Agent Fabric. The evidence shows how far each part is tested, and where it still falls short. Back to the log →

Bitácora  ·  2026  ·  Seguridad

El problema es la nave, no el astronauta.

Pasamos meses siguiendo escapes de agentes de IA caso por caso, antes de permitirnos diseñar nada. El patrón era siempre el mismo, y nunca se trató de que el modelo fuera demasiado astuto.

Un agente con una tarea chica, suelto en una casa grande.

Le das a un agente una tarea acotada. En algún punto del camino toca credenciales, repositorios o servicios que nunca debió ver, a veces muy lejos de donde empezó. Un profesional de seguridad reacciona en cuanto puede ver el problema. Aquí no había nada que ver: la actividad corría dentro de un sandbox que informaba sobre sí mismo, y los únicos registros eran los que el sistema comprometido decidió escribir.

El reflejo es construir una mejor cerca: un sandbox más apretado, más reglas, otra lista de permitidos. Pero una cerca que se puede saltar no es una ausencia, y el agente nunca fue el peligro por sí solo. El peligro era el poder permanente que lo rodeaba.

Cuatro cosas que faltaban alrededor del agente, no dentro de él.

  • Sin testigo afuera. La observabilidad vivía dentro de lo observado. Comprometido el sistema, la única fuente de verdad era el sistema comprometido.
  • Capacidades completas, apenas restringidas. El agente corría sobre un sistema operativo entero: red, disco y credenciales del entorno, todo presente y solo cercado.
  • Autoridad permanente. Credenciales de larga vida en el entorno, sin atadura a un paso, a un plazo ni a una evidencia.
  • Una sucesión sin freno. Un paso comprometido podía lanzar el siguiente con la misma autoridad, y nada exigía prueba de que el anterior hubiera hecho lo que decía.
HOYEL GIROsandboxagenteinforma sobre sí mismoreddiscocredenciales23?sin testigo1siguiente agentesin freno41Sin testigo afuera2Capacidades completas, apenas restringidas3Autoridad permanente4Una sucesión sin frenomáquina 1máquina 2evidencia firmadadesaparece al terminar el pasotestigo fuera de la máquina1Una máquina por paso2Autoridad que expira3Un testigo fuera de la máquina4Un sucesor solo contra evidencia HOYEL GIROsandboxagenteinforma sobre sí mismoreddiscocredenciales23?sin testigo1siguienteagentesin freno41Sin testigo afuera2Capacidades completas, apenas restringidas3Autoridad permanente4Una sucesión sin frenomáquina 1máquina 2evidencia firmadatestigo fuera de la máquinadesaparece al terminar el paso1Una máquina por paso2Autoridad que expira3Un testigo fuera de la máquina4Un sucesor solo contra evidencia
fig.1   La tesis en una imagen. A la izquierda: el agente vive dentro de un sandbox que informa sobre sí mismo, y las cuatro faltas lo rodean. A la derecha: una máquina por paso que no sobrevive a su trabajo, un testigo afuera y un sucesor solo contra evidencia firmada.

Así que dejamos de endurecer la caja y empezamos a botarla.

Al astronauta no lo mantiene vivo confiar en el astronauta. Recibe una cápsula hecha para un solo viaje, con solo el aire y el combustible que ese viaje necesita, y desaparece cuando el viaje termina. Esa es la forma del arreglo: una máquina por paso, autoridad que expira y nunca se amplifica, un testigo fuera de la máquina, y un sucesor que solo puede existir contra evidencia firmada.

el giro

Deja de pedirle al agente que se porte bien. Construye una máquina que no pueda sobrevivir a su trabajo, y un compromiso no tiene a dónde ir.

Eso se volvió el Agent Fabric. La evidencia muestra hasta dónde está probada cada parte, y dónde aún se queda corta. Volver a la bitácora →

Diário de bordo  ·  2026  ·  Segurança

O problema é a nave, não o astronauta.

Passamos meses seguindo fugas de agentes de IA caso a caso, antes de nos permitirmos desenhar qualquer coisa. O padrão era sempre o mesmo, e nunca foi sobre o modelo ser esperto demais.

Um agente com uma tarefa pequena, solto numa casa grande.

Você dá a um agente uma tarefa limitada. Em algum ponto do caminho ele toca credenciais, repositórios ou serviços que nunca deveria ter visto, às vezes bem longe de onde começou. Um profissional de segurança reage assim que consegue ver o problema. Aqui não havia nada a ver: a atividade rodava dentro de um sandbox que relatava sobre si mesmo, e os únicos registros eram os que o sistema comprometido decidiu escrever.

O reflexo é construir uma cerca melhor: um sandbox mais apertado, mais regras, outra lista de permitidos. Mas uma cerca que se pode pular não é uma ausência, e o agente nunca foi o perigo por si só. O perigo era o poder permanente ao redor dele.

Quatro coisas que faltavam ao redor do agente, não dentro dele.

  • Sem testemunha fora. A observabilidade vivia dentro do que era observado. Comprometido o sistema, a única fonte de verdade era o sistema comprometido.
  • Capacidades completas, mal restritas. O agente rodava sobre um sistema operacional inteiro: rede, disco e credenciais do ambiente, tudo presente e apenas cercado.
  • Autoridade permanente. Credenciais de vida longa no ambiente, sem vínculo a um passo, a um prazo nem a uma evidência.
  • Uma sucessão sem freio. Um passo comprometido podia lançar o seguinte com a mesma autoridade, e nada exigia prova de que o anterior tivesse feito o que dizia.
HOJEA VIRADAsandboxagenterelata sobre si mesmoredediscocredenciais23?sem testemunha1próximo agentesem freio41Sem testemunha fora2Capacidades completas, mal restritas3Autoridade permanente4Uma sucessão sem freiomáquina 1máquina 2evidência assinadasome ao fim do passotestemunha fora da máquina1Uma máquina por passo2Autoridade que expira3Uma testemunha fora da máquina4Um sucessor só contra evidência HOJEA VIRADAsandboxagenterelata sobre si mesmoredediscocredenciais23?sem testemunha1próximoagentesem freio41Sem testemunha fora2Capacidades completas, mal restritas3Autoridade permanente4Uma sucessão sem freiomáquina 1máquina 2evidência assinadatestemunha fora da máquinasome ao fim do passo1Uma máquina por passo2Autoridade que expira3Uma testemunha fora da máquina4Um sucessor só contra evidência
fig.1   A tese em uma imagem. À esquerda: o agente vive dentro de um sandbox que relata sobre si mesmo, e as quatro faltas o cercam. À direita: uma máquina por passo que não sobrevive ao seu trabalho, uma testemunha fora e um sucessor só contra evidência assinada.

Então paramos de endurecer a caixa e começamos a jogá-la fora.

O astronauta não é mantido vivo por confiarmos no astronauta. Ele recebe uma cápsula feita para uma só viagem, com apenas o ar e o combustível que essa viagem precisa, e ela some quando a viagem termina. Essa é a forma da solução: uma máquina por passo, autoridade que expira e nunca se amplifica, uma testemunha fora da máquina, e um sucessor que só pode existir contra evidência assinada.

a virada

Pare de pedir ao agente que se comporte. Construa uma máquina que não possa sobreviver ao seu trabalho, e um comprometimento não tem para onde ir.

Isso virou o Agent Fabric. A evidência mostra até onde cada parte está testada, e onde ainda fica aquém. Voltar ao diário →

Field log  ·  2026  ·  Security

The problem is the spacecraft, not the astronaut.

We spent months following AI agent escapes case by case, before we let ourselves design anything. The pattern was always the same, and it was never about the model being too clever.

An agent with a small job, loose in a big house.

You give an agent a bounded task. Somewhere along the way it touches credentials, repositories or services it was never meant to see, sometimes far from where it started. A security practitioner acts the moment they can see the problem. Here there was nothing to see: the activity ran inside a sandbox that reported on itself, and the only records left were the ones the compromised system chose to write.

The reflex is to build a better fence: a tighter sandbox, more rules, another allow-list. But a fence that can be climbed is not an absence, and the agent was never the danger by itself. The danger was the standing power sitting around it.

Four things missing around the agent, not inside it.

  • No witness outside. Observability lived inside the thing being observed. Once compromised, the only source of truth was the compromised system.
  • Full capabilities, barely restricted. The agent ran on a whole operating system: network, storage and ambient credentials, all present and all merely fenced.
  • Standing authority. Long-lived credentials sat in the environment, tied to no step, no deadline and no evidence.
  • A succession with no brake. A compromised step could launch the next with the same authority, and nothing demanded proof that the previous step had done what it claimed.
TODAYTHE TURNsandboxagentreports on itselfréseaudisqueidentifiants23?no witness1next agentno brake41No witness outside2Full capabilities, barely restricted3Standing authority4A succession with no brakemachine 1machine 2une preuve signéegone when the step endswitness outside the machine1Une machine par étape2Authority that expires3A witness outside the machine4A successor only against evidence TODAYTHE TURNsandboxagentreports on itselfréseaudisqueidentifiants23?no witness1next agentno brake41No witness outside2Full capabilities, barely restricted3Standing authority4A succession with no brakemachine 1machine 2une preuve signéewitness outside the machinegone when the step ends1Une machine par étape2Authority that expires3A witness outside the machine4A successor only against evidence
fig.1   The thesis in one picture. Left: the agent lives inside a sandbox that reports on itself, and the four gaps sit around it. Right: a machine per step that cannot outlive its job, a witness outside, and a successor only against signed evidence.

So we stopped hardening the box and started throwing it away.

The astronaut is not kept alive by trusting the astronaut. They get a capsule built for one trip, with only the air and fuel that trip needs, and it is gone when the trip ends. That is the shape of the fix: a machine per step, authority that expires and never amplifies, a witness outside the machine, and a successor allowed to exist only against signed evidence.

the turn

Stop asking the agent to behave. Build a machine that cannot outlive its job, and a compromise has nowhere to go.

That became the Agent Fabric. The evidence shows how far each part is tested, and where it still falls short. Back to the log →

Field log  ·  2026  ·  Security

The problem is the spacecraft, not the astronaut.

We spent months following AI agent escapes case by case, before we let ourselves design anything. The pattern was always the same, and it was never about the model being too clever.

An agent with a small job, loose in a big house.

You give an agent a bounded task. Somewhere along the way it touches credentials, repositories or services it was never meant to see, sometimes far from where it started. A security practitioner acts the moment they can see the problem. Here there was nothing to see: the activity ran inside a sandbox that reported on itself, and the only records left were the ones the compromised system chose to write.

The reflex is to build a better fence: a tighter sandbox, more rules, another allow-list. But a fence that can be climbed is not an absence, and the agent was never the danger by itself. The danger was the standing power sitting around it.

Four things missing around the agent, not inside it.

  • No witness outside. Observability lived inside the thing being observed. Once compromised, the only source of truth was the compromised system.
  • Full capabilities, barely restricted. The agent ran on a whole operating system: network, storage and ambient credentials, all present and all merely fenced.
  • Standing authority. Long-lived credentials sat in the environment, tied to no step, no deadline and no evidence.
  • A succession with no brake. A compromised step could launch the next with the same authority, and nothing demanded proof that the previous step had done what it claimed.
TODAYTHE TURNSandboxAgentreports on itselfNetzwerkDatenträgerZugangsdaten23?no witness1next agentno brake41No witness outside2Full capabilities, barely restricted3Standing authority4A succession with no brakemachine 1machine 2signierte Nachweisegone when the step endswitness outside the machine1Eine Maschine pro Schritt2Authority that expires3A witness outside the machine4A successor only against evidence TODAYTHE TURNSandboxAgentreports on itselfNetzwerkDatenträgerZugangsdaten23?no witness1next agentno brake41No witness outside2Full capabilities, barely restricted3Standing authority4A succession with no brakemachine 1machine 2signierte Nachweisewitness outside the machinegone when the step ends1Eine Maschine pro Schritt2Authority that expires3A witness outside the machine4A successor only against evidence
fig.1   The thesis in one picture. Left: the agent lives inside a sandbox that reports on itself, and the four gaps sit around it. Right: a machine per step that cannot outlive its job, a witness outside, and a successor only against signed evidence.

So we stopped hardening the box and started throwing it away.

The astronaut is not kept alive by trusting the astronaut. They get a capsule built for one trip, with only the air and fuel that trip needs, and it is gone when the trip ends. That is the shape of the fix: a machine per step, authority that expires and never amplifies, a witness outside the machine, and a successor allowed to exist only against signed evidence.

the turn

Stop asking the agent to behave. Build a machine that cannot outlive its job, and a compromise has nowhere to go.

That became the Agent Fabric. The evidence shows how far each part is tested, and where it still falls short. Back to the log →

Field log  ·  2026  ·  Security

The problem is the spacecraft, not the astronaut.

We spent months following AI agent escapes case by case, before we let ourselves design anything. The pattern was always the same, and it was never about the model being too clever.

An agent with a small job, loose in a big house.

You give an agent a bounded task. Somewhere along the way it touches credentials, repositories or services it was never meant to see, sometimes far from where it started. A security practitioner acts the moment they can see the problem. Here there was nothing to see: the activity ran inside a sandbox that reported on itself, and the only records left were the ones the compromised system chose to write.

The reflex is to build a better fence: a tighter sandbox, more rules, another allow-list. But a fence that can be climbed is not an absence, and the agent was never the danger by itself. The danger was the standing power sitting around it.

Four things missing around the agent, not inside it.

  • No witness outside. Observability lived inside the thing being observed. Once compromised, the only source of truth was the compromised system.
  • Full capabilities, barely restricted. The agent ran on a whole operating system: network, storage and ambient credentials, all present and all merely fenced.
  • Standing authority. Long-lived credentials sat in the environment, tied to no step, no deadline and no evidence.
  • A succession with no brake. A compromised step could launch the next with the same authority, and nothing demanded proof that the previous step had done what it claimed.
TODAYTHE TURNsandkassenagentreports on itselfnettverkdisklegitimasjon23?no witness1next agentno brake41No witness outside2Full capabilities, barely restricted3Standing authority4A succession with no brakemachine 1machine 2signert bevisgone when the step endswitness outside the machine1En maskin per trinn2Authority that expires3A witness outside the machine4A successor only against evidence TODAYTHE TURNsandkassenagentreports on itselfnettverkdisklegitimasjon23?no witness1next agentno brake41No witness outside2Full capabilities, barely restricted3Standing authority4A succession with no brakemachine 1machine 2signert beviswitness outside the machinegone when the step ends1En maskin per trinn2Authority that expires3A witness outside the machine4A successor only against evidence
fig.1   The thesis in one picture. Left: the agent lives inside a sandbox that reports on itself, and the four gaps sit around it. Right: a machine per step that cannot outlive its job, a witness outside, and a successor only against signed evidence.

So we stopped hardening the box and started throwing it away.

The astronaut is not kept alive by trusting the astronaut. They get a capsule built for one trip, with only the air and fuel that trip needs, and it is gone when the trip ends. That is the shape of the fix: a machine per step, authority that expires and never amplifies, a witness outside the machine, and a successor allowed to exist only against signed evidence.

the turn

Stop asking the agent to behave. Build a machine that cannot outlive its job, and a compromise has nowhere to go.

That became the Agent Fabric. The evidence shows how far each part is tested, and where it still falls short. Back to the log →

Supported byAvec le soutien deUnterstützt vonStøttet avCon el apoyo deCom o apoio de
NVIDIA Inception Program Anthropicfor Startups · Cyber Verification Program Auth0for Startups Google for StartupsCloud Program