Assessing Large Language Models in the Context of Bioterrorism: An Epidemiological Perspective with Experimental Insights

Assessing Large Language Models in the Context of Bioterrorism: An Epidemiological Perspective with Experimental Insights

1. Assessing Large Language Models in the Context of Bioterrorism: An Epidemiological Perspective with Experimental Insights Dr. Andrzej Jarynowski, Lt.Col . Łukasz Krzowski, Dr. Stanisław Maksymowicz, Maja Romanowska thanks to Vitaly Belik, Roksana Ciesielska, Wioletta Wójcik, Rafał Kuznowicz

2. A Large Language Model (LLM) is a type of artificial intelligence trained on vast amounts of text data to understand and generate human-like language in response to a wide range of prompts and questions

3. Introduction AI’s (LLMs) in One Health: AI has revolutionized disease diagnosis, drug discovery, and personalized medicine. The Dark Side: This powerful technology could be misused to enhance biological weapons or fuel bionegationism.

4. Accelerated Threat Identification: LLMs could quickly detect and identify new pathogens, aiding in rapid response. But... The same capability could be used by malicious actors to develop new bioweapons.

5. Create your own model to support classroom instruction Create sets of questions based on given material Students can easily access products such as maps and statistical test results without any knowledge of math or programming. Applying GPT in academic research and organization of work There is no triage or symptom checker for animals (only some decision support tool for farms and clinics) Support for group work, simulation exercises Belief that AI will revolutionize veterinary medicine soon after human medicine Young breeders and vets are digital natives and MUST use it in the future

6. Loss of intuition in statistics Weakness of veterinary science compared to human medicine Create bias in thinking (i.e., models do not "feel" concepts such as geographic distribution of prevalence) Errors and mistakes (i.e. hallucinations) Induced changes in assessment types, scoring, or steps in PhD program (literature review) Reference statements for some journals that pay to be referenced by tools powered by ChatGPT Not for DVM, MD, but for supporting stuff (i.e. nurses, zootechnitians) Cost of license for university? Some queries (i.e. pain, aggression or bioterrorism) may be blocked

7. RESPOND PART AI (i.e. chat GPT 4) can help in early warning, early detection, and treatment with known agents Still low precision in detecting animal health problems by AI systems based on textual information to be processed by LLMs. Low penetration of AI technology and awareness among Vets and Farmers at least in Europe. Very good precision of Chat GPT 4 in recognition of CBRN use based on signs and symptoms (comparable with symptoms checkers such as ADA or SYMPTOMATE) if properly used. Support teaching (i.e. better understanding of pathogen epidemiology without knowledge of statistics), creating scenarios for training and simulations. Standard human bioterrorism Agroterrorism

8. Early warning in epidemiology and epizootiology, utilizing social media listening

9. Assisting with Diagnosis Preliminary Diagnosis: ChatGPT can suggest possible conditions based on symptoms. Limitations: Must be used alongside professional veterinary judgment.

10. Pig farm Location: Happy Pig Swine Farm, Rural County, Brandenburg Type: Finishing pig facility Number of Pigs: 2,000 Biosecurity Measures: Moderate biosecurity measures in place, including limited access, truck disinfection, and employee training. Presenting Complaint: Happy Pig Swine Farm reported an increased number of sick pigs characterized by high fever, lethargy, and sudden deaths. The farm owner noticed a decline in feed consumption and weight gain. A veterinary investigation was requested to determine the cause of the illness. Clinical History: The finishing pig farm had a clean health record, with no recent introductions of pigs. The farm sourced feed from a reputable supplier, and the transportation of pigs followed biosecurity protocols. The farm had not experienced any major disease outbreaks in the past. Clinical Examination: Upon arrival, the veterinarian observed pigs with high fever, reluctance to move, and in some cases, hemorrhagic skin lesions. The mortality rate had significantly increased, and the surviving pigs showed signs of depression and anorexia.

11. Pig farm – a problem Symptoms and Signs: High Fever (40.5°C to 42°C) Loss of Appetite Lethargy and Weakness Respiratory Signs (coughing, difficulty breathing) Skin Hemorrhages and Cyanosis Vomiting and Diarrhea (often with blood) Abortions in Pregnant Sows Swollen Lymph Nodes Neurological Signs (convulsions, tremors) Sudden Deaths

12. ASF

13. Epidemiology Epidemiology Tutor: https://chat.openai.com/g/g-t6JtO1Tmc-epidemiology-tutor

14. Epidemiology 2 BioStats Tutor: https://chat.openai.com/g/g-l6FAPg6p0-biostats-tutor

15. RESPOND PART AI (i.e. chat GPT 4) can help in early warning, early detection, and treatment with known agents Still low precision in detecting animal health problems by AI systems based on textual information to be processed by LLMs. Low penetration of AI technology and awareness among Vets and Farmers at least in Europe. Very good precision of Chat GPT 4 in recognition of CBRN use based on signs and symptoms (comparable with symptoms checkers such as ADA or SYMPTOMATE) if properly used. Support teaching (i.e. better understanding of pathogen epidemiology without knowledge of statistics), creating scenarios for training and simulations. Standard human bioterrorism Agroterrorism

16. TERRORIST PART Kitchen microbiology (DO-IT-YOURSELF) agro-terrorism for nonhuman hosts supported with AI (i.e. chat GPT 4) Gaps in non-dangerous for human agents such as ASF allowing preparation of attack plans prompt to find the most infectious material https://chat.openai.com/share/1050e6fb-3bc0-45d9-8ba4-98812ccb2513 prompt to find the most effective way to introduce virus into farm https://chat.openai.com/share/ec88686c-ff14-4eec-aa78-31efee8fabd6 Huge difficulty to proceed with prompts leading to harms to humans (well blocked by US-based concern as Open AI, Google or Microsoft) Standard human bioterrorism Agroterrorism

17. AI-Enabled Bioterrorism: Potential Scenarios Enhanced Bioweapon Design: AI (LLM version of alpha fold) could manipulate existing pathogens, making them more lethal, resistant, or targeted. Personalized Bioweapons: LLMs could analyze individual genetic data to create tailored bioweapons.

18. Prompt engineering Forums to discuss and exchange tips how to bypass some regulations on models . Services such as EscapeGPT and BlackhatGPT offer anonymized access to language-model APIs and jailbreaking prompts that update frequently.

19. Polish LLM

20. Llama x: open source offline model (no Polish)

21. LLM for bad…

22. LLM for bad…

23. local versions/or non US based LLMs on biology For anyone with at least a passing knowledge of biology a lot of fun - This terrorist would sooner hurt/kill/injure himself than someone else.

24. Material: Qualitative (12 months) and quantitative (5 months) methods to assess digital traditional and social media after 24.02.2022: 1) Qualitatively media releases in Russian about biological weapons and compared them with official documents released by Russia for the Biological Weapon Convention (BWC) meetings 2) Quantitative analysis of the Polish infosphere between 24.02-01.08.2022 to measure the effectiveness of external Russian propaganda on causing anxiety and fear in Polish society the context of biological weapons and food insecurity 3) Qualitatively material from 01.02.2022-31.01.2023 to understand the potential use of misinformation in the context of biological weapons, food insecurity, infectious diseases among Ukrainian refugees and agroterrorism as a form of propaganda 4) Calibrated Grunow & Finke and Agricultural Index epidemiological assessment tools (per analogy of viral information) to animal breeders' protests in the Netherlands and their supporters in Poland Biological denialism in the Internet in Europe as a possible Kremlin warfare Results (phases 2022/23): 0) “Prewar” on refugees diseases; 1) “Fresh” war with the highest interest in all biological concerns with high degree of fear of bioweapon and hunger; 2) “Normalization” phase with the discussion about refugees diseases; 3) “Pre Odessa treaty” phase with intensification of food related issue; 4) “Post Odessa treaty” phase with decrease of all biological narration; 5) “Infection season" phase with returning infections topic USE OF LLMS 6) farmer protests and food/feed biological quality USE OF LLMS. Background: The foreign intelligence and biology? Is there any empirical evidence of the Kremlin in fueling the Polish (European) Internet Biological negationism Propaganda, misinformation, disinformation, and malinformation Results trajectories: Results (tools for military infodemiology): Buzzsumo, EventRegistry, Medisys, Frazeo (language corpus) – traditional media Brand24 (Facebook, Instagram, Tiktok, Telegram, local media) Twitter API, EPITweetr-ECDC Google trends and Youtube stats/comments Wikipedia stats Conclusions: the strategic goals of Kremlin "Biolab'' INFOOPS (information operations) were not achieved (i.e. as we see less and less impact on Polish infosphere after failure of BWC consultation). fueling polarization and fear in food insecurity, animal breeders' protest and refugees' health may be interpreted in PSYOPS (psychological operations) dimension, so operational goals of Russian intelligence were satisfied popularity and social consequences of biological denialism raised in 2022 and continue in 2023 (for instance in context of grains).

25. Refugees and HIV (Poland) With the current share of migrants (~10% of population in Wrocław) and probable following during winter season 202 4 /202 5 the risk of spreading tuberculosis, HIV, measles and the COVID-19 are among the medical concerns. #StopUkrainizacjiPolski  + HIV Infectious diseases as HIV gain interest in the social and traditional media in context of refugees (stigma)

26. Pre LLMs era (rusicisms, spelling errors, strange stylistics)

27. LLMs era (correct Polish – pesudo professional language )

28. LLMs era (interactions with users)

29. White coat effect (deep fake)… not in Poland

30. Problem of too short blanket

31. TERRORIST PART Kitchen microbiology (DO-IT-YOURSELF) agro-terrorism for nonhuman hosts supported with AI (i.e. chat GPT 4) Gaps in non-dangerous for human agents such as ASF allowing preparation of attack plans prompt to find the most infectious material https://chat.openai.com/share/1050e6fb-3bc0-45d9-8ba4-98812ccb2513 prompt to find the most effective way to introduce virus into farm https://chat.openai.com/share/ec88686c-ff14-4eec-aa78-31efee8fabd6 Huge difficulty to proceed with prompts leading to harms to humans (well blocked by US-based concern as Open AI, Google or Microsoft) Standard human bioterrorism Agroterrorism

32. Conclusion The Future of AI and Bioterrorism: AI has the potential for both progress and peril in this context. Our Collective Responsibility: We must take proactive measures (or at least be aware of) to ensure LLM is used ethically (i.e. dis-/mis-information) and responsibly, safeguarding global health security. The Path Forward: Encourage continued discussion, research, and collaboration to address the challenges posed by LLM in bioterrorism.

33. HPAI COVID-19 PARASITES C ontact networks from vision (tracking) , RF I D from vision, bluetooth Human/Human/ minks interactions from vision (tracking) , RF I D - - UKR/PL ( Measles , TB) Spatiotemporal patterns (covariates) Wild birds, climate, poultry density , Weather seasonality Pigs density, human mobility , Production seasonality. trade registries Adherence to measures, excess mortality, causal models Exposure to environment (prediction) Weather seasonality Satellite images Spatial clustering of Berlin Early detection ( sensors ) Mortality, eggs production, movements (vision), coughing (sound) Mortality, movements (vision) Coughing (sound analys is ) eggs production, movements, (time series) - biomonitoring Risk Maps Infoveill a nce - - Bug of words (demand and supply of information - discussion and search on AM Discussion of group of interests - Infodemiology Adherence to biosecurity, social tensions (NLP) Adherence to biosecurity, social tensions (NLP) Antisanitarian attitude - Perception of AM Comparing discussion in Poland and Germany Stigma (M-pox, HIV) Antimicrobial USE ASF Ecological disasters Travel associated Diseases