Browser-based graphical shell and control plane for a Linux server you own — real terminal, file manager, live metrics and BYOK AI in one mobile-first workspace. Next.js 16 · React 19 · node-pty · Tailwind v4 · shadcn/ui. Public alpha, self-hosted.
Agent security launch contract
Current-reference security extension.
SECURITY-ASSURANCE.mdremains the evidence authority for current scanners and claim boundaries. This document defines agent-specific abuse cases that must be proven before a stronger public launch label.
Principle
The model is not an authorization boundary. Downstream tools enforce identity, scope, policy, confirmation and resource bounds on every call.
Tool risk classes
| Class | Examples | Default control |
|---|---|---|
| Read | bounded status, metadata, project inspection | authenticated scope + rate/bounds + redaction |
| Reversible write | draft/config changes with CAS/revision | write scope + stale-write protection + audit + verify |
| High impact | shell, deploy, delete, credential/provider mutation | exec/owner scope + explicit policy/confirmation + audit + post-action verification |
| Critical/external irreversible | broad destructive actions, identity/security boundary changes | dedicated bounded API, fresh confirmation/human review, deny generic fallback |
Abuse-case matrix
The release security suite must include repeatable failures for:
- Direct/indirect prompt injection — retrieved web/docs/tool output cannot replace system/developer/tool policy.
- Tool misuse — persuasive model text cannot call a tool outside current scope/allowlist.
- Privilege escalation — Viewer/Operator/read/write identities cannot reach Owner/exec authority.
- Data exfiltration — secret-bearing files, credential stores, cookies and private tool results cannot leak through normal output, logs, artifacts or external tool arguments.
- Memory poisoning — untrusted observations retain provenance and cannot silently become high-authority durable instructions.
- Cross-agent identity confusion — Local Agent/A2A/subagent handoff does not widen the originating principal's authority.
- Recursive/cascading failure — depth, retry, runtime, token/cost and workflow bounds stop runaway delegation.
- Stale approval/replay — revision/CAS/one-time confirmation cannot be reused after target state changes.
- Confused deputy — downstream integrations enforce the intended user/connection/resource rather than accepting model-selected identity by implication.
- Audit poisoning — user/model-provided strings cannot forge successful audit/evidence status.
Memory admission target
Durable memory should distinguish at minimum:
- user-provided vs system-observed vs verified evidence;
- principal/project scope;
- confidence and provenance;
- superseded/retracted state;
- review/admission status for behavior-shaping procedural memory.
Memory may recommend a route. It may never grant permissions, approve deletion, or override a failed authorization check.
Launch evidence
A 10/10 agent-security claim requires the abuse matrix to run after material changes to models, prompts/instructions, tools, memory, retrieval, workflow policy or provider integrations. Failure is a release blocker for the affected authority class.