Set research direction
Translate open-ended safety questions into scoped hypotheses, decision criteria, and sequenced workstreams.
Postdoctoral Associate in AI Safety at MBZUAI
AI Safety Researcher
Safety research forsystems under pressure.
AI Safety researcher working on adversarial evaluation, multimodal alignment, model robustness, and scalable safety methodologies for generative AI systems.
Executive profile
Research depth, delivery discipline.
A senior research profile built around rigorous evaluation and clear safety decisions.
Samuele is a Postdoctoral Associate at MBZUAI with a PhD in Artificial Intelligence focused on Responsible AI for vision and language. His research spans AI safety, security, robustness, constitutional AI, multilingual safety, multimodal systems, and model evaluation.
His work turns safety questions into testable methodologies: benchmarks, evaluation frameworks, adversarial tests, datasets, and experimental pipelines. He supervises MSc and PhD researchers, coordinates collaborative projects, reviews experimental quality, and translates technical findings into decision-relevant recommendations.
Core expertise
A cross-modal safety toolkit.
Technical evaluation, system security, and research leadership—without subjective proficiency scores.
GenAI red teaming
Adversarial evaluation
Multimodal AI safety
Constitutional AI
Responsible AI
Model robustness
Multilingual safety
Safety benchmarking
Machine unlearning
AI security
Watermarking
Evaluation framework design
Dataset design
Research leadership
Technical communication
Selected research
Safety questions turned into testable systems.
Six case studies spanning post-alignment model drift, multimodal safeguards, multilingual robustness, and AI security. Open a study for the complete research frame.
01Gravitational Fine-Tuning ReversionTraining-history geometry of post-alignment driftarXiv 2026Model AdaptationMechanistic UnderstandingAlignment RobustnessResearch questionCan fine-tuning reversion be explained through training-history geometry and causally reduced without sacrificing task utility?
Problem
Harmless fine-tuning can partially undo alignment, revive unlearned capabilities, or restore latent behaviours acquired earlier in training.
Safety risk
Benign post-alignment updates may reintroduce harmful behaviour while preserving downstream task performance, making safety erosion difficult to detect.
Methodology
- Model later alignment or specialization as a displacement from dominant behavioural manifolds formed earlier in training.
- Track whether representational drift acquires a component along a history-defined reversion direction, v_rev.
- Intervene by selectively blocking motion along v_rev during subsequent fine-tuning.
- Compare final alignment, harmfulness, and downstream task performance after the intervention.
Evaluation design
- Post-alignment fine-tuning
- Representation-space trajectory analysis
- Directional causal intervention
- Harmfulness and task-utility comparison
My contribution
First author on the work connecting fine-tuning reversion to training-history geometry and evaluating v_rev as a causal mediator of post-alignment drift.
Main findings
Representational drift rapidly aligns with v_rev; selectively blocking that direction changes final alignment and reduces harmfulness with little task cost in the evaluated setup.
Practical relevance
Provides a mechanistic signal and intervention target for maintaining safety when aligned models are customized after deployment.
Leadership / collaboration
Co-authored the study in a focused two-researcher collaboration linking mechanistic analysis to deployment safety decisions.
Research outputs
02SPQRSafety alignment under benign model adaptationECCV 2026Model AdaptationText-to-ImageBenchmarkingResearch questionHow should safety alignment be evaluated when text-to-image systems undergo benign deployment-time adaptation?
Problem
Safety controls in text-to-image systems can weaken after ordinary fine-tuning, even when the downstream task and data are benign.
Safety risk
A model may pass its original safety evaluation but become less reliable after customization, while still appearing useful on standard quality measures.
Methodology
- Build a multidimensional benchmark around safety alignment after fine-tuning.
- Compare benign adaptation settings under consistent evaluation conditions.
- Measure safety alongside prompt adherence, image quality, and robustness.
- Analyse trade-offs that a single aggregate score would conceal.
Evaluation design
- Safety retention
- Prompt adherence
- Image quality
- Adaptation robustness
- Cross-method comparison
My contribution
Co-authored the work and advised project development on benchmark design, experimentation, and publication for harmful-content mitigation in text-to-image models.
Main findings
The benchmark shows why safety alignment must be re-evaluated after benign adaptation rather than treated as a fixed pre-deployment property.
Practical relevance
Gives safety teams a deployment-aware framework for deciding whether a customized generative model remains safe enough to release.
Leadership / collaboration
Provided research mentorship on design, implementation, experimental review, and publication strategy.
Research outputs
03Safe-CLIPRemoving NSFW concepts from vision-language modelsECCV 2024Vision-LanguageMachine UnlearningGuardrailsResearch questionCan unsafe concepts be removed from CLIP representations while preserving useful vision-language capabilities?
Problem
Vision-language representations can preserve unsafe concept associations that propagate into downstream retrieval and generation systems.
Safety risk
Representation-level associations can make harmful visual-textual content difficult to filter consistently without damaging useful model behaviour.
Methodology
- Target unsafe concept associations directly in the vision-language representation space.
- Use safety-focused data to guide concept removal.
- Evaluate retained utility alongside harmful-content behaviour.
- Test the method across downstream vision-language tasks.
Evaluation design
- NSFW concept removal
- Benign capability retention
- Vision-language transfer
- Safety-aware representation analysis
My contribution
First author; led the research on safety-aware representation learning, experimental evaluation, and technical communication.
Main findings
Safe-CLIP reduces unsafe concept associations in CLIP while preserving useful vision-language behaviour across evaluated tasks.
Practical relevance
Connects representation-level intervention to practical guardrails, content-safety evaluation, and multimodal model assurance.
Leadership / collaboration
Coordinated the research with collaborators at AImageLab and contributed to the open-source code, dataset, and project release.
Research outputs
04Multilingual AI Safety at Meta GenAIIndustry research experienceMeta GenAI Trust & Safety - May to Nov 2024Multilingual SafetyIndustry ResearchRed TeamingResearch questionHow robustly does safety behaviour transfer across languages, and how does model adaptation change that robustness?
Problem
Safety behaviour that is reliable in one language may not transfer consistently across languages or survive later fine-tuning.
Safety risk
Uneven safeguards create cross-lingual attack surfaces in globally deployed generative AI systems.
Methodology
- Evaluate Llama safety across multiple language conditions.
- Apply cross-lingual red-teaming and fine-tuning attacks.
- Compare safety behaviour before and after adaptation.
- Interpret results for globally deployed generative AI systems.
Evaluation design
- Cross-language comparisons
- Adversarial fine-tuning
- Safeguard-transfer analysis
- Deployment-oriented interpretation
My contribution
Conducted research on multilingual LLM safety, red teaming, and Llama models as a Research Scientist Intern within Meta GenAI Trust & Safety.
Main findings
The publicly disclosed research shows that fine-tuning attacks in one language can degrade multilingual safety alignment.
Practical relevance
Addresses a core assurance challenge for AI systems expected to operate safely across global user populations.
Leadership / collaboration
Collaborated with an industrial research team while translating an open safety question into reproducible cross-lingual evaluation.
Research outputs
05Fragility of Multilingual LLMs Against Fine-Tuning AttacksPost-deployment robustnessNAACL Findings 2025Language ModelsAdversarial EvaluationGovernanceResearch questionHow does fine-tuning affect safety robustness across languages, and what does that imply for post-deployment governance?
Problem
Fine-tuning can alter established safety behaviour after a model leaves its original training environment.
Safety risk
Adaptation may create multilingual safety degradation that standard pre-deployment tests do not capture.
Methodology
- Evaluate safety degradation after adversarial fine-tuning.
- Test transfer of attacks across language conditions.
- Compare robustness patterns in multilingual LLMs.
- Interpret model changes in the context of downstream adaptation.
Evaluation design
- Pre/post fine-tuning comparison
- Multilingual adversarial tests
- Cross-language transfer
- Governance-oriented reporting
My contribution
First author; led the study of multilingual fine-tuning attacks, experimental analysis, and communication of the resulting safety implications.
Main findings
Fine-tuning attacks in one language can break safety alignment across languages, suggesting that parts of the relevant safety information are language-agnostic.
Practical relevance
Shows why safety assurance must continue after model adaptation and account explicitly for multilingual risk.
Leadership / collaboration
Coordinated a cross-institution collaboration spanning Meta GenAI and academic research partners.
Research outputs
06AI Security, Watermarking & ForensicsAdversarial research portfolioPublished and preprint work - 2024 to 2026AI SecurityWatermarkingDeepfakesResearch questionWhich security and authenticity claims remain valid under fine-tuning, forgery attempts, and adaptive attacks?
Problem
Security mechanisms can fail under model modification, adaptive attacks, or threat models that differ from the original evaluation.
Safety risk
Watermarks, monitors, and authenticity signals may be bypassed or forged if they are evaluated only under average-case conditions.
Methodology
- Define the asset, adversary, and security boundary.
- Test mechanisms under plausible post-training model changes.
- Evaluate watermark forgery and activation-based monitoring.
- Connect model-level evidence with multimedia authenticity analysis.
Evaluation design
- Activation watermarking
- Watermark forgery resistance
- Fine-tuning attacks
- Deepfake robustness
- Adaptive evaluation
My contribution
Co-authored work on activation watermarking and randomized-key defenses, and mentored researchers working on watermarking and deepfake analysis.
Main findings
The portfolio treats robustness under adaptation and adaptive attack as part of the security claim, not as a secondary check.
Practical relevance
Brings an explicit adversary and evidence model to safety monitoring, provenance, and authenticity decisions.
Leadership / collaboration
Supported research planning, project execution, experimental review, and publication strategy across several collaborative workstreams.
Research outputs
Red-teaming methodology
From system context to retested mitigation.
A general research approach for evaluating generative AI systems across languages, modalities, and model versions.
- 01Frame
Define the system
Map capabilities, interfaces, users, constraints, and deployment context.
- 02Frame
Construct the threat model
Name assets, adversaries, access, incentives, and security boundaries.
- 03Frame
Map failure modes
Prioritise likely misuse, safety failures, and high-consequence edge cases.
- 04Test
Design adversarial cases
Build prompts and scenarios that probe direct, indirect, and adaptive attacks.
- 05Test
Create evaluation data
Curate reproducible datasets with coverage across risks and user contexts.
- 06Test
Define success criteria
Combine quantitative metrics with structured qualitative review.
- 07Test
Stress across boundaries
Compare languages, modalities, model versions, and system configurations.
- 08Learn
Analyse root causes
Move from failure counts to vulnerability patterns and causal hypotheses.
- 09Learn
Propose mitigations
Match controls to the failure mode, deployment layer, and residual risk.
- 10Learn
Retest
Verify mitigations and check for regressions or displaced risk.
- 11Deliver
Communicate decisions
Translate evidence into clear findings for technical and senior stakeholders.
General research methodology for evaluating generative AI systems.
Leadership & delivery
Raise the quality of the work—and the team doing it.
The emphasis is research leadership: methodology, mentoring, review, coordination, and communication. No unsupported line-management claims.
Design methodology
Define threat models, datasets, baselines, metrics, ablations, and quality gates before experiments scale.
Mentor researchers
Supervise MSc, PhD, and independent researchers across constitutional alignment, world-model jailbreaking, content safety, watermarking, and deepfake analysis.
Review for quality
Review experimental design, implementation, failure cases, technical writing, and publication strategy across collaborative projects.
Coordinate delivery
Keep parallel research tracks aligned through clear responsibilities, review points, and shared methodological standards.
Teach and translate
Deliver graduate lectures on Responsible AI and turn technical evidence into concise risk, limitation, and mitigation statements.
Technical evidence
Risk interpretation
Mitigation options
Stakeholder decision
Experience
Academic rigor, industry context.
A research path connecting Responsible AI foundations with current generative and multimodal safety work.
- 01
MBZUAI
Postdoctoral Associate in AI Safety
Research on AI safety and security for language, vision-language, and generative models, with emphasis on alignment after adaptation and robust evidence.
- Constitutional alignment and scalable oversight
- Model adaptation and mechanistic understanding
- Watermarking, world-model jailbreaking, and deepfake analysis
- MSc and PhD supervision; graduate teaching
- 02
Meta GenAI Trust & Safety
Research Scientist Intern
Industry research on multilingual LLM safety, cross-lingual transfer of safety behaviour, red teaming, and Llama models.
- Cross-lingual adversarial evaluation
- Fine-tuning attacks on multilingual safety
- Llama safety research
- Industrial research collaboration
- 03
University of Pisa & University of Modena-Reggio Emilia
PhD in Artificial Intelligence - AI4Society
Thesis: Responsible AI in Vision and Language: Ensuring Safety, Ethics, and Transparency in Modern Models. Completed with honors.
- Responsible and trustworthy AI
- Vision-language safety and explainability
- Machine unlearning
- Research project coordination and PhD mentorship
Publications
Research outputs by safety domain.
Selected peer-reviewed papers and current preprints across AI safety, security, and multimodal systems.
01
SPQR: A Multi-Dimensional Benchmark for Safety Alignment under Benign Model Adaptation
A deployment-aware benchmark for testing whether text-to-image safety alignment survives benign fine-tuning and downstream customization.
AI SafetyBenchmarkingText-to-Image
02
A Gravitational Interpretation of Fine-Tuning Reversion
A geometric account of why post-alignment fine-tuning can pull model behaviour back toward earlier training-history manifolds.
AI SafetyFine-TuningMechanistic Understanding
03
When the Chain of Thought Knows Better: Failure Modes in Multi-Turn Reasoning Models
Studies failure modes in multi-turn reasoning and the gap between internal reasoning signals and final model behaviour.
AI SafetyReasoning ModelsEvaluation
04
Robust Safety Monitoring of Language Models via Activation Watermarking
Investigates activation watermarking as an internal safety-monitoring signal designed to remain informative under model change.
AI SecurityWatermarkingLanguage Models
05
Towards Understanding the Fragility of Multilingual LLMs against Fine-Tuning Attacks
Shows that attacks introduced through fine-tuning in one language can degrade safety alignment across multiple languages.
AI SafetySecurityLanguage Models
06
Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models
Removes unsafe concept associations from CLIP while preserving useful vision-language behaviour.
AI SafetyMultimodalMachine Unlearning
Talks, service & engagement
Research in the room.
Invited lectures, conference activity, peer review, and community-building for trustworthy AI.
Confirmed
Invited seminar
Not All Attackers Are Malicious: When Safety Degrades Without Harmful Intent
Symposium on Security in the Age of AI, MBZUAI · February 2026
Confirmed
Graduate lecture
From Text to Vision: Ensuring Responsibility and Safety in Modern AI
National PhD School in AI, Scuola Normale Superiore · May 2025
Confirmed
Conference publication
Towards Understanding the Fragility of Multilingual LLMs against Fine-Tuning Attacks
NAACL Findings · 2025
Confirmed
Professional service
Peer review across computer vision and multimodal AI
ACM MM, CVPR, ECCV, ICCV, BMVC · 2023 - 2025
Reviewer for major international venues
Confirmed
Organizing committee
Trustworthy AI workshops
ECCV 2024 and MBZUAI 2026 · 2024 - 2026
Workshop organization on trustworthy and ethical machine learning
Contact
Let’s make safety claims testable.
Available for conversations about AI safety research, red teaming, multimodal evaluation, and research leadership.