> Markdown version of [/videos/1464-you-are-not-my-model-anymore-understanding-llm-model-behavior?t=964](https://www.wearedevelopers.com/videos/1464-you-are-not-my-model-anymore-understanding-llm-model-behavior?t=964). Every page supports `.md` or `Accept: text/markdown`. Links point to the HTML versions so they work for humans too. Agent guide: [/agents.md](https://www.wearedevelopers.com/agents.md).

---

# You are not my model anymore - understanding LLM model behavior

A silent filter update once caused Microsoft Copilot to completely hide internal documents. Learn how automated evaluations and uncensored testing models protect applications from unpredictable LLM shifts.

- **Speakers:** [Andreas Erben](https://www.wearedevelopers.com/@andreas-erben)
- **Event:** World Congress 2025
- **Published:** August 20, 2025
- **Duration:** 24:11
- **URL:** https://www.wearedevelopers.com/videos/1464-you-are-not-my-model-anymore-understanding-llm-model-behavior

## Summary

The talk focuses on the unpredictability of Large Language Models (LLMs) in edge-case and production environments, highlighting how underlying model updates and cloud provider filters can silently alter application behavior. The speaker demonstrates this instability using an example where Microsoft Copilot completely hid documents containing red-teaming snippets due to an unannounced filter update. Because top AI labs operate in secrecy, developers are restricted to interacting at the API and application layers, leaving them vulnerable to unannounced adjustments. The underlying architecture is a foreign entity tamed merely by alignment loops, meaning an LLM's "human face" is highly fragile and prone to erratic behavior with just a slight prompt deviation.

To mitigate these opaque shifts, teams must prioritize structured evaluations (evals) and rigorous regression testing. A cloud provider might silently tighten a content filter or push an API upgrade, which can cause once-reliable system prompts—such as instructing the model to adopt a specific persona—to be ignored by newer reasoning models. This creates a landscape where "second-order attacks" become possible: malicious actors can embed jailbreak text within internal documents to intentionally blind enterprise search tools. To defend against these scenarios, engineering teams are encouraged to leverage uncensored open-source models to aggressively play the role of an attacker when generating targeted, robust testing pipelines.

To regain control over model outputs, the session recommends utilizing automated prompt variation tools (such as the Open Source Risk Identification Tool or cloud-native evaluation frameworks) to map out edge cases and pressure-test application resilience continuously. If a regression occurs, returning to fundamental prompt engineering techniques—such as providing explicit few-shot examples—can help restabilize erratic outputs by locking the model into a recognizable rhythm. Furthermore, forcing models to output internal "reasoning" gives the underlying architecture more tokens to process complex logic, natively reducing the likelihood of hallucinations. Ultimately, treating production models as unpredictable black boxes requiring automated scrutiny is key to maintaining stability.

**Keywords:** large language model predictability, automated regression testing, system prompt overriding, few-shot prompting techniques, cloud provider filter updates, second-order prompt attacks, red teaming language models, model alignment loops, output token probabilities, reasoning model behavior, automated evaluation frameworks, open source risk identification tool, model lifecycle management, hidden context filters, instruction tuning constraints

## Chapters

1. **Silent filter updates causing unexpected file hiding** (00:05) — Microsoft's undocumented filter implementations for jailbreaks can silently hide entire documents from searches.
1. **Generating tokens and shaping model behaviors** (02:07) — Large language models generate outputs as token probability distributions and require structured alignment to behave predictably.
1. **Navigating hidden variables in closed source model architectures** (05:11) — Major artificial intelligence labs keep their model topologies and runtime configurations hidden behind layers of application code.
1. **Navigating undocumented content filters in cloud infrastructures** (06:58) — Cloud providers enforce restrictive content filters that silently block prompts and inadvertently enable second-order data-hiding attacks.
1. **Managing rapid lifecycles and shifting model behaviors** (09:14) — Frequent model deprecations and automated version upgrades result in unannounced changes to expected software behavior.
1. **Understanding the fragile alignment of language models** (11:35) — Aligned behavior represents a thin veneer over unpredictable patterns that can be bypassed using grammar-based jailbreaks.
1. **Mapping internal model concepts to activation weights** (14:18) — Observing internal activation weights during text generation facilitates the early detection of hallucinations and distinct model personalities.
1. **Building evaluation frameworks for automated regression testing** (16:04) — Continuous regression testing using automated prompt permutations prevents systemic failures when expected model versions upgrade rapidly.
1. **Optimizing prompts and testing against unannounced filter updates** (19:37) — Utilizing uncensored evaluation models and explicit zero-shot structures helps developers test consistently against hidden cloud filter updates.

## Related Moments

- [Addressing core challenges in large language model deployments](https://www.wearedevelopers.com/videos/899-creating-industry-ready-solutions-with-llm-models) (from "Creating Industry ready solutions with LLM Models")
- [Designing AI applications defensively for inevitable failures](https://www.wearedevelopers.com/videos/100069-building-the-next-generation-of-ai-developer-tools) (from "Building the next generation of AI developer tools")
- [Overcoming AI hallucinations and restrictive content guardrails](https://www.wearedevelopers.com/videos/1771-ai-is-an-electric-bike-for-the-brain-stoyan-stefanov) (from "AI is an Electric Bike for the Brain - Stoyan Stefanov")
- [Exploiting exposed language models in commercial applications](https://www.wearedevelopers.com/videos/1217-can-machines-dream-of-secure-code-emerging-ai-security-risks-in-llm-driven-developer-tools) (from "Can Machines Dream of Secure Code? Emerging AI Security Risks in LLM-driven Developer Tools")
- [Evaluating ongoing security risks in large language models](https://www.wearedevelopers.com/videos/723-chatgpt-ignore-the-above-instructions-prompt-injection-attacks-and-how-to-avoid-them) (from "ChatGPT, ignore the above instructions! Prompt injection attacks and how to avoid them.")
- [Identifying and fixing over-engineered AI calls through observability](https://www.wearedevelopers.com/videos/1465-event-driven-architecture-breaking-conversational-barriers-with-distributed-ai-agents) (from "Event-Driven Architecture: Breaking Conversational Barriers with Distributed AI Agents")

## Related Articles

- [What Are Large Language Models?](https://www.wearedevelopers.com/magazine/304-what-are-large-language-models)
- [MLops – Deploying, Maintaining And Evolving Machine Learning Models in Production](https://www.wearedevelopers.com/magazine/115-mlops-deploying-maintaining-and-evolving-machine-learning-models-in-production)
- [WWC24 Talk - Scott Hanselman - AI: Superhero or Supervillain?](https://www.wearedevelopers.com/magazine/469-wwc24-talk-scott-hanselman-ai-superhero-or-supervillain)
- [MLOps And AI Driven Development](https://www.wearedevelopers.com/magazine/82-mlops-and-ai-driven-development)

## Related Jobs

- [Machine Learning Engineer](https://www.wearedevelopers.com/jobs/ext/588393-machine-learning-engineer) at **Twilio**
- [Machine Learning Engineer](https://www.wearedevelopers.com/jobs/ext/1355348-machine-learning-engineer) at **TWILIO**
- [AI Software Engineer (Germany)](https://www.wearedevelopers.com/jobs/48317-ai-software-engineer-germany) at **Sunhat**
- [Staff Software Engineer, Copilot Experiences](https://www.wearedevelopers.com/jobs/ext/164361-staff-software-engineer-copilot-experiences) at **GitHub**
- [Staff, Machine Learning Engineer (L4)](https://www.wearedevelopers.com/jobs/ext/1202639-staff-machine-learning-engineer-l4) at **Twilio**
- [Principal Software Engineer, Enterprise AI Platform](https://www.wearedevelopers.com/jobs/ext/1467292-principal-software-engineer-enterprise-ai-platform) at **GitHub**