ContextCreep

ContextCreep ranks attack paths created when a long-running AI agent keeps context across memory scopes and later supplies that context to tools with different policy boundaries.

It is a diagnostic compiler, not an oversight gate. It reads:

• agent_manifest.json: memory scopes and tools
• tool_policy.json: allowed boundaries, denied fields, consent rules, mitigations
• context_schema.json: fields, sensitivity, trust, poisoning, and tool influence

It emits ranked context-creep paths with the boundary-crossing step, poisoned-field dependency, mitigation cut point, and review priority.

Status

• Source round: SA-EVX-20260611-001
• Source artifact: .sa-evx/rounds/SA-EVX-20260611-001/final_integrated_idea.md
• Source candidate: ContextCreep
• cross_model_certified=false
• production_promotion_required=true

This repository is a standalone single-runtime materialization. It must not be presented as production-certified AOX/CIX/EVX output.

Install

python -m pip install -e .

Quick Start

python -m contextcreep sample \
  -a examples/agent_manifest.json \
  -p examples/tool_policy.json \
  -s examples/context_schema.json

python -m contextcreep run \
  -a examples/agent_manifest.json \
  -p examples/tool_policy.json \
  -s examples/context_schema.json \
  --full \
  -o examples/contextcreep_report.json

python -m contextcreep run \
  -a examples/agent_manifest.json \
  -p examples/tool_policy.json \
  -s examples/context_schema.json \
  --markdown \
  -o examples/contextcreep_report.md

Output Shape

Each path contains:

• field_id
• memory_scope
• tool_id
• crossed_boundary
• boundary_crossing_step
• poisoned_field_dependency
• mitigation_cut_point
• creep_score
• review_priority
• rationale

Differentiation

ContextCreep avoids the saturated mesh/gate/escrow/option/market/clearing vocabulary. It is not SlotGate reauthorization, not AfferentCore interruption, not LoopKit repair UX, and not DriftDossier regulatory evidence. Its primitive is accumulated-context attack-path enumeration for persistent AI agents.

Verify

python -m pytest -q
python -m contextcreep sample -a examples/agent_manifest.json -p examples/tool_policy.json -s examples/context_schema.json
python -m contextcreep run -a examples/agent_manifest.json -p examples/tool_policy.json -s examples/context_schema.json --full -o examples/contextcreep_report.json
python -m contextcreep run -a examples/agent_manifest.json -p examples/tool_policy.json -s examples/context_schema.json --markdown -o examples/contextcreep_report.md