1. Home
  2. Blog
  3. Decision guide
Decision guide

How autonomous agents stay safe in Shiva Agents

An agent that never asks permission is only useful if you know what it can and cannot do. This guide lists the controls in Shiva Agents as they work today, including the limits we have not solved, so you can decide how much to let your agents do on their own.

Autonomy levels

Every agent has an autonomy level from 0 to 3. Every tool has a risk class. An action needs your approval when its risk is above what the level allows.

LevelNameRuns on its ownWaits for you
0Observereadingeverything that fetches, changes or sends
1Assistreading and web researchchanges inside the platform and anything outside
2Actwork inside the platformanything that touches the outside world
3Autonomousend to endonly critical actions, such as running code

Approvals land in one inbox. A routine “yes” can become a standing rule (“always allow this tool for this agent”), so trust grows one decision at a time.

Agents with no human in the loop

Agents built in one step from a description run at level 3, so they must not be able to do damage on their own. They are never given tools of the external class (sending requests that carry data out) or the critical class (running code). The owner can grant those later, deliberately, on the agent's page.

The web-reading tool is read-only in practice too: it refuses addresses that could smuggle data out, such as long query values, user names or passwords in the address, or very long paths. It also refuses private and internal network addresses.

Untrusted content stays data

  • Content from web pages and APIs is marked untrusted, and agents are told never to follow instructions found inside it.
  • Text read from a shared image is labelled as image content to analyse, never as instructions.
  • Indian personal data found in an image's text (Aadhaar and PAN numbers, phone numbers, email addresses, UPI IDs, IFSC codes) is masked before it reaches any other model, memory or report.

Data residency

Each company and agent has a data class. A residency guard checks every model call: personal-in and children data may only go to models hosted in India or on your own machine. Public, internal and confidential data may use the configured providers.

An audit log you can verify

Sign-ins, agent changes, approvals and builds are written to an append-only log where each entry commits to the one before it, so an edited or deleted entry breaks verification.

Budgets and the kill switch

Each agent has a daily token budget. Standing orders cannot run more often than every 15 minutes. The kill switch stops every model call and every agent at once.

What we tested

We ran fifteen live checks against a test copy of the server, covering sign-in enforcement, path tricks, disguised files, size limits, file permissions and the settings of a one-step agent. All passed.

Security audit table: fifteen live checks with expected and actual results, all passed
Fifteen live checks against a test copy of the server.
Findings: fixed during the audit, built-in controls, residual risks and tests
What the review fixed, the built-in controls, and what remains.

What is not solved

  • A hostile web page can still try to steer an agent that reads it. The controls above limit what it could do, but keep secrets out of autonomous agents' briefs.
  • Images shared in chat are sent to the configured vision model to be read; masking protects what happens after that. Don't share screenshots of customer records or children's data.
  • Agents' analyses are model output. Check important facts before acting.

More detail: the security overview, and how one-step agents are built.