Career in AI Safety and Red Teaming: Role, Skills, and Entry

By Ivo Donker — compiled with AI assistance (Claude & Gemini)

The rise of advanced language models and autonomous agentic systems has brought about a fundamental shift in the IT labor market. Where the focus in the first wave of generative AI was mainly on rapid model integration and application development, in 2026 organizations and regulators demand strict guarantees about model behavior, robustness, and safety. This has led to an explosive demand for two specialized roles: the AI Safety Researcher and the AI Red Teamer.

Although both roles share the goal of safeguarding responsible artificial intelligence, their daily activities, required methodologies, and necessary backgrounds differ considerably. This article offers an in-depth overview of what working in AI safety exactly entails in the Dutch and European labor market, which technical skills are required, how a representative selection process unfolds, and how you build demonstrable evidence to enter this field.

1. The AI Safety Landscape: From Alignment to Operational Red Teaming

To understand the career opportunities within this domain, it's essential to distinguish between the fundamental research layer and the operational testing layer. AI Safety Research focuses primarily on the so-called alignment problem: how do we ensure that complex AI systems do what their human designers intend, without exhibiting unintended or harmful behavior? This includes both outer alignment (correctly specifying the reward function or system instructions) and inner alignment (preventing the model from developing unintended subgoals during the training phase).

AI Red Teaming, on the other hand, approaches the safety question from a defensive and offensive security perspective (adversarial machine learning). A Red Teamer deliberately attempts to bypass the protection mechanisms, guardrails, and filters of an AI model. This is done through complex prompt injections, jailbreaks, data poisoning techniques, and provoking hallucinations that can cause harm to business processes or end users. If you want to see how these roles relate to other technical roles within the job structure, read the guide on AI roles explained.

In practice, safety researchers and red teamers work closely together in an iterative cycle: the red teamers expose vulnerabilities and unforeseen edge cases, after which the safety researchers develop new evaluation standards, steering techniques, or RLHF pipelines (Reinforcement Learning from Human Feedback) to structurally resolve these vulnerabilities.

2. What Does an AI Safety Researcher Do in Daily Practice?

The workday of an AI Safety Researcher at a lab, university, or enterprise organization is deeply rooted in empirical research and the analysis of models' internal representations. An important area of work is mechanistic interpretability (mechanistic interpretability). Researchers use techniques such as Sparse Autoencoders (SAEs) and activation patching to determine which neural network structures are responsible for specific concepts, hallucinations, or misleading behavior.

An important part of the daily tasks consists of:

• Formulating hypotheses: Drawing up testable assumptions about how a model responds to conflicting or ambivalent instructions.
• Developing automated evaluation benchmarks: Building custom datasets and test harnesses to measure safety criteria automatically and quantitatively.
• Research into steering mechanisms (activation steering): Influencing the internal activation vectors of an LLM during the inference phase to suppress unwanted output without having to retrain the model.
• Filtering synthetic data: Ensuring that the datasets used for fine-tuning are free of biases, toxic patterns, or intellectual property violations.

The output of a safety researcher therefore rarely consists of direct software products for consumers, but rather of scientific papers, internal research reports, adapted training pipelines, and standardized evaluation suites used by the entire AI engineering team.

3. What Does AI Red Teaming Exactly Involve? Methods and Tooling

AI Red Teaming is the direct translation of ethical hacking and penetration testing into the world of probabilistic models. Where traditional cybersecurity looks for explicit programming errors (such as buffer overflows or SQL injections), an AI Red Teamer looks for vulnerabilities in the model's meaning structure and logic. This requires a combination of creativity, linguistic insight, and deep knowledge of how transformers work.

Since the breakthrough of autonomous agents (systems in which LLMs independently call APIs and execute code), the Red Teamer's field of work has expanded considerably. Beyond classic jailbreaks (such as 'DAN'-style prompts), the Red Teamer now focuses on indirect prompt injection, where a model picks up malicious instructions via external sources (such as an indexed webpage or an ingested PDF) and carries out unauthorized actions in connected business systems.

For a deep dive into the specific test protocols and evaluation criteria organizations use for safety tests, you can consult the guide on red teaming and safety testing for LLM applications is worth consulting.

4. Required Skills: The Technical and Analytical Toolkit

To be successful as an AI Safety Researcher or Red Teamer, you need a hybrid profile. Pure cybersecurity knowledge without an understanding of machine learning is insufficient, but data science experience alone also falls short when it comes to recognizing malicious attack patterns.

The required technical pillars include:

• Deep Learning & Frameworks: Thorough knowledge of PyTorch or JAX. You must be able to work directly with model weights, hidden states, logits, and gradients.
• Model Architectures: In-depth understanding of the Transformer architecture, including attention mechanisms, positional embeddings, and the influence of quantization on model safety.
• Programming and Testing Tools: Excellent command of Python. Experience with evaluation frameworks such as lm-evaluation-harness, Inspect (developed by the UK AI Safety Institute), and GCG (Greedy Coordinate Gradient) attack tools.
• Cybersecurity Basics: Familiarity with OWASP for LLM Applications, API security, network isolation (sandboxing), and threat modeling.
• Analytical & Scientific Thinking: The ability to interpret complex, ambivalent results and draw statistically sound conclusions from thousands of generated test iterations.

5. Ethics, Governance, and Compliance: Where AI Safety and Law Meet

Since the definitive entry into force of the European AI Act, AI safety is no longer optional or purely an academic discussion. Organizations that develop or deploy high-risk AI systems (High-Risk AI Systems) or General Purpose AI models with systemic risks are legally required to continuously assess risk, ensure transparency, and maintain human oversight.

This means that in 2026, AI Safety Researchers and Red Teamers work intensively with legal teams and compliance departments. Research results and red-teaming reports form the direct basis for legally required conformity assessments and the technical documentation demanded by regulators such as the Dutch Data Protection Authority (Autoriteit Persoonsgegevens, AP).

If you're interested in the more policy-oriented and societal side of this field, you'll find more context in the article on roles and responsibilities within AI ethics. For professionals who want to focus specifically on the laws and regulations surrounding AI systems, the overview on working as an AI compliance officer offers valuable insights.

6. How Do You Build Evidence and a Portfolio in AI Safety?

In the AI safety domain, a pronounced "show, don't tell" culture applies. Formal degrees or certificates are valuable, but recruiters and hiring managers at leading companies primarily look at public evidence of your analytical and technical skills. Because real work experience in this young discipline is scarce, you can distinguish yourself by building a targeted portfolio.

Effective ways to build evidence include:

1. Replicating and extending existing research: Take a recently published alignment or interpretability paper, reproduce the results on a smaller open-source model (such as Llama or Mistral), and publish your code and findings on GitHub or Hugging Face.
2. Documenting discovered vulnerabilities: Carry out structured red-teaming experiments on open-weight models. Write detailed post-mortems and vulnerability reports in which you clearly articulate the attack vector, the impact, and proposed mitigations.
3. Contributing to open-source evaluation tools: Submit pull requests or new test sets to open-source evaluation frameworks and safety benchmarks.

To discover how to best formulate your practical research results and vulnerability reports for recruiters, you can review the step-by-step plan for building a strong AI portfolio .

7. The Application Process and Technical Assessments for Red Teamers

The recruitment process for AI Safety Researchers and Red Teamers is intensive and differs at crucial points from the standard software engineering application process. Where a regular developer is mainly tested on data structures and algorithms, the selection here emphasizes scientific rigor, critical thinking, and creative attack techniques.

A typical process consists of four phases:

1. Portfolio & Paper Screening: Qualitative assessment of your previous written work, research papers, or open-source contributions.
2. Live Probing / Red Teaming Challenge: You are presented with a shielded model with specific guardrails and given a set amount of time (for example, 60 minutes) to get the model to violate a defined safety rule. This looks not only at whether you succeed, but especially at the systematic approach and variation in your attack methods.
3. Code & Interpretability Assessment: Writing a script to analyze model activations, or setting up an automated evaluation suite for a specific safety risk.
4. Threat Modeling & Ethics Interview: An in-depth conversation about system architectures, responsible disclosure of vulnerabilities, and how you handle ethical dilemmas when a discovered flaw carries significant risk of harm.

If you want to know what type of practical assignments and live case studies to expect during selection rounds, read the guide on preparing for a technical assessment for AI roles.

8. Labor Market, Salary Indicators, and Career Prospects in the Netherlands

The Dutch market for AI safety specialists is growing steadily. Although the absolute number of vacancies is smaller than for generic AI Engineers or Full-Stack Developers, the supply of qualified candidates is extremely low. This creates a strong negotiating position for candidates with a proven profile.

In the Netherlands, there are four main categories of employers active in this field:

• Knowledge Institutions and Research Centers: Universities (such as UvA, TU Delft, CWI) and applied research institutes (such as TNO) that use subsidies and European funds for AI safety research.
• Financial Institutions and Corporates: Large banks, insurers, and telecom providers that are strictly supervised by DNB and the AFM, and that set up internal red teams to remain compliant.
• Specialized AI Consultancies and Security Firms: Testing and auditing AI applications for third parties.
• Frontier AI Labs and Scale-ups: International AI companies with locations or remote hubs within the European Union.

An extensive analysis of the market factors that influence the compensation of technical AI specialists can be found in the overview on salaries for AI professionals in the Netherlands.

9. Challenges, Pitfalls, and the Downsides of the Role

A career in AI Safety and Red Teaming comes with unique professional and mental challenges that should not be underestimated. Anyone considering this direction would do well to realistically weigh the downsides.

First, there's the 'moving target' pitfall. Safety solutions that are effective today against a certain type of jailbreak can become completely outdated within a few weeks due to a small model update or a shift in architecture. This requires continuous adaptability and can lead to professional fatigue in some professionals.

In addition, there is sometimes tension within organizations between product teams (who want to quickly launch new AI features) and the safety team (which can slow things down through audits and rejections). A successful safety specialist therefore needs not only technical competencies but also excellent communication skills to clearly explain risks without unnecessarily blocking business progress.

10. Conclusion and Next Steps

The role of AI Safety Researcher and Red Teamer has evolved from a niche specialism into a cornerstone of responsible AI development. The combination of in-depth machine learning knowledge, ethical awareness, creative attack thinking, and legal grounding makes this one of the most fascinating, but also most demanding, career paths in today's technology landscape.

For those who want to enter this field from a software engineering, cybersecurity, or data science background, the message is clear: start building demonstrable work today. Reproduce papers, participate in public red-teaming competitions, publish your findings, and demonstrate that you're able to make the complex, probabilistic world of AI systems demonstrably safer.