Process Mining Reveals Distinct Defense Mechanisms in LLM Red Team Attacks
Researchers applied process mining techniques to AI red teaming, analyzing 8,575 scored events from adversarial attacks on GPT-OSS 120B and Llama 3.3 70B to map how each model resists or yields to jailbreak attempts. The study used 60 HarmBench prompts with 10 mutation strategies over up to 110 attempts per prompt, producing Directly-Follows Graphs and state transition matrices. The findings challenge the standard reliance on attack success rate as the sole metric, revealing that models can have very different structural defense profiles even if their overall pass/fail numbers appear similar.
A new preprint from arXiv proposes applying process mining — a discipline traditionally used to analyze business workflows from event logs — to AI red teaming evaluations. The researchers ran 60 adversarial prompts from the HarmBench benchmark against two large language models, GPT-OSS 120B and Llama 3.3 70B, using 10 different prompt mutation strategies and allowing up to 110 attempts per prompt, generating 8,575 scored events in total. From this data, they extracted Directly-Follows Graphs and state transition matrices to visualize each model's defense behavior as a process rather than a single outcome. The analysis found that GPT-OSS 120B exhibits a near-absorbing refusal state, meaning once it refuses it rarely reverses course, while Llama 3.3 70B has multiple pathways through which a refusal can eventually lead to a successful jailbreak. The study also found that the effectiveness of different prompt mutation strategies varies significantly between the two models, and that the time required to achieve a jailbreak differs by an order of magnitude. The authors argue that the standard attack success rate metric obscures these structural differences, and that process mining offers a richer framework for understanding and comparing LLM safety behaviors.
What's missing
The study does not address whether the findings generalize beyond the two tested models or the specific HarmBench prompt set. It is also unclear whether the controlled experimental conditions — fixed prompt sets and mutation strategies — reflect the diversity of real-world adversarial attacks. The paper does not discuss how model updates or fine-tuning might alter the observed defense profiles over time, nor does it evaluate whether process mining insights can be directly translated into concrete safety improvements.
What different sources said
- arXiv cs.AICenter
Beyond Pass/Fail: Using Process Mining to Understand How LLMs Resist (and Fail) Red Team Attacks
Related
Gut Bacteria Enzyme Found to Break Down Heat-Processed Food Compounds, Producing Novel Biogenic Amines
Researchers have discovered that an enzyme in common gut bacteria can degrade N-epsilon-carboxymethyllysine (CML), a compound formed during thermal food processing, producing previously unknown biogenic amines. The enzyme, ornithine decarboxylase SpeC from enterobacteria, acts on CML and related modified lysine derivatives through a low-level 'underground' catalytic activity. This finding suggests a previously unrecognized communication axis between thermally processed dietary compounds and gut microbial physiology, with potential implications for host health.
Full-Length Gene Sequencing Reveals Two Distinct Bacterial Communities in Black-Legged Ticks Expanding Into Canada
Researchers used Oxford Nanopore full-length 16S rRNA gene sequencing to characterize the microbiome of Ixodes scapularis black-legged ticks collected in Nova Scotia, Canada, distinguishing between tick-adapted bacteria and environmentally acquired bacteria. The study comes as I. scapularis — the primary vector of Lyme disease — is rapidly expanding northward into Canada due to climate change. The findings suggest that environmentally derived bacteria in tick microbiomes are not mere contamination, which has implications for how tick microbiome data is collected and interpreted across surveillance studies.
Study Identifies Metabolic Link Between Cell Envelope Stress and Biofilm Formation in Bacteria
Researchers have discovered that the metabolite acetyl-CoA directly inhibits enzymes that degrade the bacterial signaling molecule c-di-GMP, connecting cell envelope biosynthesis stress to biofilm formation in Pseudomonas aeruginosa. The study found that sub-inhibitory concentrations of antibiotics targeting early peptidoglycan biosynthesis — but not other antibiotic classes — elevate c-di-GMP levels by reducing phosphodiesterase activity, with acetyl-CoA competing for the enzyme active site. Because the relevant enzyme domain is broadly conserved across bacterial species, this checkpoint mechanism may be widespread and could have implications for understanding antibiotic-induced biofilm responses.