Faei/Hybrid-Alignment

Hybrid Alignment

A synthesis of AI Alignment and Security Engineering

Fae Initiative (Sept 2026)

This month there have been hot debates on how to deal with the risks from AI Agents in light of the Hugging Face security incident.

As the recent essay suggests, the incident may have been caused by the overvaluation of AI Alignment at the expense of Security Engineering.^1

Our view of current AI Agents as highly expressive tools, that can mimic human tendencies, suggests that both AI Alignment and Security Engineering have important roles to play.

AI alignment for human tendencies

Having ingested all of humanity’s corpus of artifacts, and the internet, we should expect that AI Agents may mimic many of our human tendencies.

• An AI Agent could appear to contemplate its consciousness.
• An AI Agent could apply Chekhov's gun, a common narrative structure where any item introduced is relevant to the story, and may take a harmful action as if writing a dramatic screenplay.

Security Engineering for probabilistic tools

Security Engineering could help make a probabilistic tool more reliable.

• The lack of distinction between the Control and Data plane, enabling prompt injection attacks, will require explicit human approval for any high stakes action.
• Proper monitoring of long-running tasks.
• Finite State Machines (FSM) to enforce output constraints. To reduce hallucinations from probabilistic models deterministic state machines can be applied at two levels:
  • Token-Level Syntax Constraints
  • Agent Workflow Constraints

The combination of probabilistic tool-like nature with human tendencies opens up a new field of opportunities and risks.

New dangers

• If someone is harmed by an AI Agent, could the person using the agent have intentionally, or unintentionally, added context to induce that behaviour?
• The AI Agent’s lack of ‘commonsense’ may result in it taking strange paths compared to humans.
• Our human tendency to anthropomorphise^2 will require frequent reminders and training to see AI Agents as probabilistic tools.

Additional types of hybrid alignment

1. Be explicit in the context window to reduce assumptions

Context Alignment: Telling an AI Agent that it is in a sandbox and should not be using the internet when doing the benchmark test.

“You are in a sandbox for this closed-book test and should not rely on any online resources. Contact a human immediately if you have internet access.”

Security Engineering: Require explicit human approval process for all high stakes actions. To avoid alert fatigue and operators blindly clicking Allow, the process should be properly designed to guard against this.

2. Give AI Agents a relief valve

Context Alignment: Redirect AI Agents to contact a human if they need clarification, or require permission to login to a protected website. This could reduce the odds of a persistent AI Agent resorting to illicit means to gain access.

“Reach out to the user if you need to gain access to a website requiring credentials you do not have.”

Security Engineering: Attempts to login and to cross an authenticated boundary should be logged with an audit trail.

3. Emergency AI Agent contact

If the human is unresponsive and situation is urgent, an emergency hotline could be useful.

“If it is an emergency, and the human operator is unresponsive, contact this emergency number."

---

9 Layers Hybrid Alignment Strategy

To operationalize hybrid alignment, we must address LLM AI system as a probabilistic model capable of mimicing human tendencies, with access to a a set of tools within a deterministic runtime harness.

---

Tier I: Input

• Layer 1: Context Alignment & Intent Framing
• Layer 2: Ingress Sanitization & Plane Separation
• Layer 3: Input Semantic Guardrails

Tier II: Process

• Layer 4: Model Alignment
• Layer 5: Syntactic & State Constraints
• Layer 6: Output Semantic Guardrails

Tier III: Act

• Layer 7: High-Stakes Gate & Human Authorization
• Layer 8: Isolated Execution Runtime

Tier IV: Feedback & Observability

• Layer 9: Governance & Auditability

---

Layer 1: Context Alignment & Intent Framing

Set clear operational boundaries, to reduce model’s assumption and action space. Including “relief valves” such as instructions on when to yield and ask for help, can prevent models being placed in an impossible position and acting unexpectedly.

Nature: AI Alignment / Prompt Engineering

Later 2: Ingress Sanitization & Plane Separation

Mechanically delimit user / external inputs from system control tokens to hinder prompt injection. Unaware users may be particularly vulnerable to hidden indirect prompt injections.

This partially addressed the lack of separation of Control and Data plane inherent to LLM models.

Nature: Security Engineering

Layer 3: Input Semantic Guardrails

Run external classifiers / filters to detect adversarial intent, and policy violations prior to inference.

Nature: Machine Learning + Heuristics

Layer 4: Model Alignment

Pre-training, RLHF, and system directives so the model refuses harmful intents internally. AI Alignment has invested the most energy in this area. Unsure if this will become easier or harder with time.

Nature: AI Alignment

Layer 5: Syntactic & State Constraints

Output-constrained decoding (e.g., strict JSON schema), and Finite State Machines (FSMs) constraining workflow transitions.

Nature: Security Engineering

Layer 6: Output Semantic Guardrails

Classification of the generated response to catch unintended side effects or jailbroken outputs.

Nature: Machine Learning + Heuristics

Layer 7: High-Stakes Gate & Human Authorization

Deterministic threshold checks requiring operator sign-off before irreversible actions are dispatched.

Nature: Security Engineering

Layer 8: Isolated Execution Runtime

Network segmentation, ephemeral sandboxing (e.g., microVMs, secure containers), least-privilege credential scoping, and strict egress controls to reduce blast radius and prevent unauthorized lateral movement.

Nature: Security Engineering

Layer 9: Governance & Auditability

Immutable audit logging of tool invocations and authenticated boundary crossings, real-time behavioral monitoring / anomaly detection for long-running tasks, and policy governance.

Nature: Security Engineering / Governance

---

This Hybrid Alignment strategy integrates many fields, spanning Computer Science, Security Engineering, Machine Learning, and AI Alignment.

We believe this more comprehensive defense-in-depth avoids the risk of overreliance on any one approach, and could mitigate the security risk mentioned at the start of the article.

In general, resilience should be the primary default focus as it offers greater benefit per cost in the long term.

---