For years, AI safety has been framed as a problem of better training. Better data. Better objectives. Better alignment techniques. Better reinforcement learning. Better fine tuning. Better prompting.
But across an increasingly diverse body of research, a quieter conclusion has emerged.
Training alone cannot guarantee safe behavior in deployed AI systems.
Not because researchers are careless. Not because alignment is unimportant.
But because learned systems are probabilistic, adaptive, opaque, and fragile under distribution shift. These properties are intrinsic. No amount of optimization removes them.
As a result, safety is being rediscovered not as a property of models, but as a property of systems.
And systems require control planes.
This post synthesizes a growing body of work across alignment, robustness, formal methods, governance, and deployment research that all point to the same architectural conclusion:
AI safety requires an explicit, external, runtime enforcement layer.
In other words: Middleware.
We built SASKI. It is a built and tested safety middleware that already implements this architecture in real deployed conversational systems.
01Training Is Not Enforcement
One of the clearest empirical signals comes from work analyzing how safety training fails in practice.
Research on jailbreaks, adversarial prompting, and distribution shift shows that even models trained with extensive safety objectives can be induced to violate constraints at inference time. The failure mode is not rare edge cases. It is structural.
Safety training produces preferences, not guarantees.
A trained model may prefer to refuse harmful content, but when the input distribution changes, when context accumulates over a session, or when multiple objectives compete, those preferences can be overridden.
This is not a flaw in the training process. It is a consequence of probabilistic generalization.
Once a model is deployed, there is no theorem that says its behavior remains within the intended safety envelope. At best, there is hope backed by empirical testing.
Hope is not a safety mechanism.
02Inference Time Is Where Risk Lives
Several recent papers make an important shift in emphasis. They stop asking how to train safer models and start asking how to monitor and control behavior while the model is running.
This includes work on:
- Inference time safety classifiers
- Latent space steering vectors
- Continuous evaluation during generation
- Runtime output monitoring
- Forced termination or refusal mechanisms
The shared insight is simple:
Most dangerous failures do not occur at the initial prompt. They occur mid response, after context accumulation, or at decision boundaries where the model is uncertain.
This is why safety mechanisms that only operate during training or at prompt time systematically miss real failures. (Note: Hybrid approaches combining internal steering with external middleware can further enhance robustness, as explored in related works on latent knowledge discovery.)
Once again, the implication is architectural. If safety decisions must be made during inference, then safety cannot live solely inside static model weights.
It must live alongside execution.
03Explicit Safety Signals Beat Implicit Alignment
Another thread of research critiques the idea that safety should emerge implicitly from a single objective function.
Instead, these papers argue that safety must be represented as an explicit signal with authority over generation.
This includes:
- Binary safety classifiers that dominate decoding
- Dedicated safety reasoning pathways
- Hard constraints that override helpfulness or fluency
- Deterministic refusal insertion
- Session level safety state tracking
When safety is treated as an implicit preference, it competes with other preferences.
When treated as an explicit signal, it governs them.
This distinction matters.
In traditional engineering, we do not encode “do not explode” as a soft objective. We implement pressure limits, kill switches, and structural constraints.
AI systems are finally being treated with the same seriousness.
04Formal Methods Point to External Control
Work in formal verification and control theory offers a parallel insight.
Research on reward machines, formal specifications, and verifiable abstractions emphasizes the separation of:
- Task logic from learning
- Control flow from optimization
- Specification from execution
These frameworks exist precisely because learned systems are difficult to reason about directly.
Rather than trying to prove properties of a neural network, formal methods introduce an external structure that constrains what the network is allowed to do. (For instance, reward machines support temporal logic for specifying extended behaviors like sequences and conditionals.)
The lesson carries over cleanly to modern AI systems. If you want guarantees, you do not place them inside an opaque, adaptive function approximator. You place them in a layer that can be inspected, audited, and enforced.
That layer is not the model.
05Governance Research Already Assumes Middleware
Policy and governance oriented research makes this explicit, even if it does not use the word middleware.
System cards, preparedness frameworks, and risk management documents consistently describe safety architectures that include:
- Input and output filters
- Runtime monitoring tools
- Deterministic policy enforcement
- Logging and audit trails
- Kill switches and escalation paths
- Human override mechanisms
- These are not training techniques. They are system components.
Importantly, they are described as separate from the model itself. The model is treated as one component in a larger pipeline, not as the final authority.
This is an admission, not a weakness.
It reflects the reality that deployment decisions and safety responsibilities live outside the training loop.
06The Seldonian Insight: Guarantees Require Separation
The Seldonian framework adds a final piece to the puzzle.
It argues that if you want probabilistic guarantees about behavior, you must:
- Specify constraints explicitly
- Verify them on held out data
- Accept or reject behavior based on confidence thresholds
- Crucially, the Seldonian approach separates optimization from verification.
Learning proposes. Verification disposes.
This is exactly the pattern emerging in deployed AI systems.
The model generates. The safety layer decides whether the output is allowed.
Once again, the architecture is clear.
07The Convergence Pattern
Across all of this research, from wildly different domains, the same conclusions recur:
- Learned models are powerful but unreliable
- Safety preferences are not safety guarantees
- Distribution shift is inevitable
- Opaqueness cannot be eliminated
- Verification must be external
- Enforcement must be deterministic
- Monitoring must occur at runtime
- Auditability is non optional
No single paper declares “middleware is the future.”
They do not need to.
They are all independently rebuilding parts of it.
- Some do it inside the model with special tokens and steering vectors.
- Some do it with external classifiers.
- Some do it with formal abstractions.
- Some do it with governance frameworks.
But structurally, they are converging on the same answer.
08Why Middleware Is Inevitable
Middleware solves a problem that no training method can.
It allows us to:
- Treat safety as a first class system property
- Enforce hard constraints without retraining
- Operate across multiple models and vendors
- Adapt policies without touching model weights
- Log decisions for regulators and insurers
- Fail closed instead of failing silently
- Upgrade safety independently of capability
This is not a rejection of alignment research. It is its completion.
Training makes models useful. Middleware makes them governable.
09The Future Architecture of AI Safety
The emerging architecture looks less like a single aligned model and more like this:
- A probabilistic model optimized for capability
- A deterministic safety layer that monitors inputs and outputs
- Explicit policy definitions
- Runtime evaluation and state tracking
- Hard enforcement actions
- Audit logs and explanations
- Human escalation paths
This is not speculative. It is already how serious systems are being built.
What is changing is that researchers are now publishing the reasons why this architecture is necessary.
10The Quiet Shift
There will be no single paper that declares the end of model only safety.
Instead, there is a quiet shift underway.
Alignment research is becoming more honest about its limits. Deployment research is becoming more explicit about enforcement. Governance frameworks are assuming external control. Formal methods are re entering the conversation.
The result is not a new algorithm.
It is a new layer.
Middleware is not a workaround.
It is the natural endpoint of taking AI safety seriously.
And the fact that this conclusion is being reached independently, across disciplines, is the strongest evidence that it is correct.
\-- Across alignment research, formal methods, governance frameworks, and deployment studies, a consistent conclusion is emerging: training shapes behavior distributions, but safety requires explicit, external, runtime enforcement. Middleware is not an optional add-on. It is the architectural layer where guarantees, auditability, and governance become possible
Our SASKI system is a built and tested safety middleware that operationalizes this convergence today, rather than treating it as a future research problem.
References and Further Reading
01Safety failures under training-only approaches
Jailbreaks and inference-time failuresKang, D. et al. Jailbroken: How Does LLM Safety Training Fail? arXiv:2307.02483
Demonstrates systematic bypass of training-time safety via adversarial prompting and distribution shift.
Wei, J. et al. Jailbreak Attacks Against LLMs. arXiv:2307.08715
Shows that safety behaviors learned during training degrade predictably under novel contexts.
Key takeaway: training induces preferences, not guarantees.
02Explicit safety signals and runtime dominance
Li, J. & Kim, J.-E. Safety Alignment Can Be Not Superficial With Explicit Safety Signals. arXiv preprint.
Argues that safety must be represented as an explicit signal with authority over generation, not a soft objective.
OpenAI. o1 System Card. 2024
Documents inference-time monitoring, policy enforcement, and forced termination mechanisms.
OpenAI. GPT-4o System Card. 2024
Describes system-level safety mitigations at input and output boundaries, including deterministic enforcement.
Key takeaway: safety must dominate decoding, not compete with it.
03Inference-time monitoring and latent separation
Li, Y. et al. Steer LLM Latents for Hallucination Detection. ICML 2025
Introduces Truthfulness Separator Vectors (TSV) to reshape representations at inference time without retraining.
Burns, C. et al. Discovering Latent Knowledge in Language Models Without Supervision. arXiv:2212.03827
Shows that models internally represent truth but do not reliably act on it without external intervention.
Key takeaway: detection and enforcement must occur during execution.
04Formal methods and separation of specification from learning
Icarte, R. T. et al. Reward Machines: Exploiting Reward Function Structure in Reinforcement Learning. JAIR 2022
Separates task logic from learning, enabling modularity, interpretability, and verifiable control.
McIlraith, S., et al. Reward Machines and Temporal Logic for Control.
Demonstrates how structured specifications can govern learned agents externally.
Key takeaway: learned policies should operate inside externally defined constraints.
05Constraint-based guarantees and verification
Thomas, P. et al. The Seldonian Framework. JMLR 2019
Introduces algorithms that enforce probabilistic constraints with high-confidence guarantees.
Hoag, A. et al. The Seldonian Toolkit. IEEE/ACM ICSE-Companion 2023
Introduces algorithms that enforce probabilistic constraints with high-confidence guarantees.
Key takeaway: learning proposes, verification disposes.
06Governance, preparedness, and system-level safety
Amodei, D. et al. Concrete Problems in AI Safety. arXiv:1606.06565
Early articulation of the need for monitoring, interruption, and containment.
Bengio, Y. et al. Managing Extreme AI Risks Amid Rapid Progress. Science, 2024
Argues for enforceable system-level safety mechanisms, safety cases, and runtime controls.
OpenAI. Preparedness Framework. 2023–2024
Establishes continuous evaluation, kill switches, and escalation paths outside model training.
Key takeaway: responsibility and enforcement live outside the model.
07Foundational critiques of probabilistic safety
Madry, A. et al. On Evaluating Adversarial Robustness. ICLR 2019
Shows why best-effort defenses without guarantees fail under adaptive pressure.
Goodfellow, I. et al. Explaining and Harnessing Adversarial Examples. ICLR 2015
Demonstrates intrinsic fragility of learned systems under small perturbations.
Key takeaway: robustness must be engineered, not hoped for.
Learn more about our SASKI technology HERE
Test our SASKI middleware in our sandbox app at www.mytrusted.ai
.