Sandbox Escapes: Isolation Boundaries as a Design Input

Monthly research note. Theme: Adversarial Infrastructure & Global Systems.

TL;DR

A focused memo on Sandbox Escapes: Isolation Boundaries as a Design Input: define the model, state the properties, then design the system so those properties remain true under failure and adversaries.

Key takeaways

• Evidence pipelines (audit/config history) are part of incident response correctness.
• Protect observability: you can’t respond blind, and telemetry can be attacked.
• Dependencies (DNS, routing, PKI) are shared attack surfaces—plan containment.
• Write assumptions down; treat them as interfaces.
• Make failure modes explicit and observable.

Why this matters

• Incident response is a protocol: practice it, automate it, validate it.
• Logs are only useful if they remain trustworthy under compromise.
• Attackers exploit cost asymmetry: make abuse cheap and defense expensive.
• Privacy failures often come from metadata, not plaintext.

Key questions

• How do you detect attacks that look like “normal traffic spikes”?
• Where is the attacker’s leverage (routing, DNS, dependency, identity, time)?
• How do you make abuse expensive (proof-of-work, quotas, pricing, friction)?
• Which controls fail first under load: auth, rate limits, storage, or observability?
• What is your degraded-mode behavior (and is it safe)?
• Which logs are trustworthy under compromise (append-only, signed, isolated)?

Assumptions

• Attackers can manipulate routing and DNS indirectly (upstream failures, BGP issues).
• Traffic spikes can be malicious or accidental; you must handle both.
• Observability pipelines can be attacked (cardinality explosions, log injection).
• Operators are human and will make mistakes under pressure.

Non-goals

• Assuming WAF/rate limits are sufficient without architecture changes.
• Relying on dashboards that vanish during the incident.

Model & invariants

Defense is about cost asymmetry. If the attacker spends $1$ and you spend $100$, you lose.

Treat observability as a dependency: protect it from overload and manipulation.

Define which operations fail closed vs fail open. Do it before an incident.

Security properties

• Downgrade resistance: negotiation can’t silently weaken security posture.
• Evidence: critical actions emit verifiable audit events.
• Authenticity: actions are bound to identity and purpose.
• Replay resistance: duplicated inputs do not change outcomes.

Failure modes

• Config drift that weakens security posture over time.
• Observability gaps during incidents (missing evidence).
• Recovery paths that only work when nothing is broken.
• Mixed-version behavior that violates assumptions silently.

Design sketch

flowchart TD
  edge["Edge (rate limits + WAF)"] --> core["Core Services"]
  core --> data["Data Plane"]
  data --> control["Control Plane"]
  control --> edge
  siem["Detection/Response"] --> core
  siem --> edge

Implementation notes

Keep evidence pipelines alive: you can’t respond blind.

Evidence checklist:
- Immutable logs (append-only)
- Signed audit events
- Time sync monitoring
- Dependency health snapshots
- Config change history

Verification strategy

• Incident replay: reconstruct timeline from evidence pipelines.
• Dependency chaos: DNS issues, cert failures, upstream outages.
• Game days: simulate DDoS, dependency failure, and credential abuse.
• Observability stress: cardinality explosions and sampling under attack.
• Policy tests: fail closed/open behaviors are unit-tested.

Operational notes

• Document and rehearse degraded-mode policy with on-call rotations.
• Instrument cost: which defenses become expensive and when.
• Make emergency controls quick: feature flags, circuit breakers, safe defaults.
• Protect the edge and the evidence: rate limits + SIEM + log integrity.
• Keep recovery paths simple: restore from known-good, rotate secrets, reissue certs.

What to monitor

• Authz failures and policy denials (unexpected spikes).
• Rollback events and the conditions that triggered them.
• Error budget burn + tail latency under load.
• Admission-control / rate-limit rejections (by reason).
• Invariant violation rate (should be ~0).

Rollback plan

• Prefer backward-compatible changes; avoid “flag day” upgrades.
• Keep dual-write / dual-verify windows where appropriate.
• Define an explicit rollback trigger (metrics + thresholds).
• Preserve evidence (configs, artifacts, audit logs) to reconstruct what changed.
• Use canaries and staged rollout; stop early when signals degrade.

Evidence

• Learn TLA+ (1) — Practical entry point for specification and model checking.
  • Evidence: Model the smallest thing that can break; use model checking to validate invariants before optimizing.
• Designing Data-Intensive Applications (Kleppmann) (2) — The systems-engineering baseline for correctness, replication, and failure.
  • Evidence: Replication and consistency tradeoffs as engineering constraints; use as reference when naming guarantees.

Open questions

• How do you keep control-plane access during widespread incidents?
• Where do you pay cost asymmetry today—and can you flip it?
• What is your ‘safe mode’ when dependencies fail?
• Which operation, if abused, causes irreversible damage?

Checklist

• Telemetry captures correctness signals.
• Failure modes enumerated with mitigations.
• Safety properties stated as invariants.
• Rollback plan rehearsed and automated.
• Costs bounded (CPU/memory/bandwidth) under adversarial inputs.
• Assumptions listed and reviewed.

Further reading

• RFC 4271: BGP-4 — Routing is part of your threat model whether you like it or not.
• Let's Encrypt Incident Reports — Operational failures and recovery in real-world PKI.
• RFC 6480: An Infrastructure to Support Secure Internet Routing — RPKI basics and why routing security is hard operationally.
• Cloudflare Outage (July 2, 2019) Postmortem — A concrete example of global failure, containment, and recovery lessons.
• Learn TLA+ — Practical entry point for specification and model checking.
• Designing Data-Intensive Applications (Kleppmann) — The systems-engineering baseline for correctness, replication, and failure.