Ikarus AI Labs01SafeAI
Agent Authority
What can this AI agent potentially do?
- identity
- principal
- delegation
- tools
- MCP servers
- access modes
- data reachability
- unknown authority
Ikarus AI Labs · Research programme
Evidence-Based Assurance for AI-Mediated Software Change
Connecting agent authority, software ecosystem intelligence, and accountable human oversight.
AI is becoming part of the software change process. We study how organisations can make those changes observable, explainable and accountable through machine-readable evidence.
Central question
What evidence should exist before an AI-mediated software change is allowed to proceed?
§1 The problem
Software change is becoming AI-mediated
Traditional assurance assumes
- developers make changes
- humans understand the change
- dependencies are relatively passive
- permissions are relatively stable
Agents can independently
- Discover
- Select
- Modify
- Configure
- Execute
Assurance must therefore connect
- Who
- Can do what
- Changed what
- Depends on what
- What evidence exists
- What should happen
§2 The evidence model
Four evidence planes
02OpenPulse
Software Ecosystem Intelligence
What is happening to the software the agent is changing or selecting?
- lifecycle
- support
- upstream changes
- vulnerabilities
- ownership
- applicability
- evidence confidence
03Policy
Assurance Decision
Given the available evidence, what should happen?
- ALLOW
- REVIEW
- BLOCK
- ACCEPTED EXCEPTION
04Record
Accountability
Why was this decision made, and under whose authority?
- evidence
- policy
- baseline
- approver
- exception
- timestamp
- provenance
Hypothesis This is an active research programme, not a solved problem.
§3 Research instruments
Instrument A · Agent Authority Evidence
SafeAI
An open-source static analyser for understanding the authority an AI agent can potentially exercise.
Discover
What can it access?
Compare
What changed?
Govern
Should it be allowed?
Unknown is an evidence state, not evidence of safety.
Instrument B · Software Ecosystem Intelligence
OpenPulse
An evidence-based intelligence system for understanding meaningful changes in the open-source software ecosystem.
- RELATED ≠ AFFECTED
- UNKNOWN ≠ NOT AFFECTED
- MATCH STRENGTH ≠ EVIDENCE STRENGTH
- SOURCE SIGNAL ≠ ACTIONABLE
What changed upstream, what evidence establishes it, and does it apply to us?
§4 Why both
Two evidence streams. One assurance question.
SafeAI and OpenPulse are intentionally separate research instruments. The research question is whether combining them produces better decisions about AI-mediated software change.
HYPOTHESIS · not a claim the problem is solved
SafeAI
What authority does the agent have?
OpenPulse
What is happening to the ecosystem it touches?
Assurance policy
AllowReviewBlock
§5 Principles
Our research principles
- P1Evidence over assertion
- Every meaningful conclusion should have supporting evidence.
- P2Provenance matters
- Evidence should retain where it came from and how it was interpreted.
- P3Unknown is first-class
- Unknown must never silently become safe.
- P4Applicability matters
- A finding is not automatically an impact.
- P5Deterministic where possible
- Policy decisions should be reproducible and explainable.
- P6Human accountability
- Automation should strengthen accountability rather than erase it.
- P7Open research
- Where possible, datasets, benchmarks, schemas and tools should be openly inspectable.
- Source
- Evidence
- Interpretation
- Policy
- Decision
- Accountability
Build the evidence layer for AI-mediated software
AI will increasingly participate in software decisions. The challenge is not to prevent every autonomous action, but to understand what authority exists, what changed, what evidence supports the decision, and where human accountability remains necessary.