arXiv:2606.08270v1 [cs.CR] 6 Jun 2026
An AI Security Agent for University ACMIS: Multi-Vector Threat Detection and Automated Response Joseph Walusimbi
Joshua Benjamin Ssentongo
Dept. of Electronics and Computer Engineering Soroti University Soroti, Uganda [email protected]
Dept. of Electronics and Computer Engineering Soroti University Soroti, Uganda [email protected]
Abstract—University Academic Management Information Systems (ACMIS) are high-value targets for a wide spectrum of security threats including brute-force login attacks, payment fraud, privilege escalation, insider data theft, and academic integrity violations. Traditional rule-based intrusion detection systems are inadequate because many malicious activities are structurally indistinguishable from normal operations. This paper presents an AI-based security agent for ACMIS that combines supervised anomaly detection, behavioural analytics, and a natural language processing chatbot for secure password recovery. The agent monitors five operational layers: authentication, authorisation, financial transactions, user behaviour, and system health, and responds through a four-tier risk escalation framework. A modular architecture allows the core engine to be extended to other institutional systems. Experiments on a simulated ACMIS event log dataset demonstrate a threat detection macro-average F1 of 0.91, compared to 0.49 for a rule-based baseline, with critical-tier automated response latency under 300 ms at the 95th percentile. Index Terms—academic management systems, anomaly detection, behavioural analytics, brute-force detection, insider threat, AI security agent, chatbot, password recovery, ACMIS, university cybersecurity
I. I NTRODUCTION University information systems occupy an unusual threat profile: they manage sensitive personal data alongside highstakes academic operations, yet are typically operated with security postures more appropriate to general administrative systems than mission-critical infrastructure [1]. An Academic Management Information System (ACMIS) aggregates all of these functions into a single platform, making it a concentrated target for both external attackers and malicious insiders. Critically, many of the most damaging attacks against ACMIS exploit behavioural gaps: actions that individually appear legitimate but collectively constitute fraud, data theft, or system abuse. A lecturer whose account gradually accumulates administrative privileges; a student who submits a payment that is later reversed; a script that generates Payment Registration Numbers (PRNs) at machine speed—each looks unremarkable in isolation but is detectable through behavioural analysis.
Existing intrusion detection systems (IDS) rely predominantly on signature-based rules, which are effective against known attack patterns but fundamentally limited against novel or slow-moving threats [2]. Machine-learning anomaly detection has been studied extensively in network security [3] and financial fraud [4], but its systematic application to university management systems—with their unique combination of academic, financial, and identity management functions—remains underexplored. This paper makes the following contributions: • A multi-layer AI security agent covering five operational surfaces of ACMIS: authentication, authorisation, financial transactions, user behaviour, and system health. • A four-tier automated response framework escalating from passive monitoring to emergency lockdown based on a calibrated risk score. • An NLP-powered chatbot for intelligent, fraud-resistant password recovery that detects mass reset attacks during the recovery flow. • A modular architecture enabling the core anomaly engine to be adapted to banking, healthcare, and other institutional systems with sector-specific plug-in modules. • An evaluation on simulated ACMIS event logs demonstrating significant improvement over a rule-based IDS baseline. The remainder of this paper is organised as follows: Section II reviews related work; Section III characterises the threat model; Section IV describes the architecture; Section V details detection and response; Section VI presents results; Section VII discusses deployment and limitations; and Section VIII concludes. II. R ELATED W ORK A. Anomaly Detection in Information Systems Chandola et al. [3] provide the canonical taxonomy of anomaly detection, distinguishing point, contextual, and collective anomalies. ACMIS threats span all three: a PRN generated 500 times (point), a login at 3 a.m. for a user whose baseline is
8 a.m.–5 p.m. (contextual), and a slow data exfiltration pattern spanning weeks (collective). This breadth necessitates a multimodel approach. B. Machine Learning for Intrusion Detection Buczak and Guven [2] surveyed machine learning approaches for network intrusion detection, finding that ensemble methods and deep learning consistently outperformed singlemodel approaches. More recently, recurrent neural network language models have demonstrated strong performance on user-entity behaviour analytics tasks [8], making sequential models natural candidates for modelling the temporal structure of ACMIS user sessions. C. Financial Fraud Detection Ahmed et al. [4] reviewed fraud detection for financial transactions, identifying velocity, ratio, and recency features from temporal sequences as particularly discriminative. These principles transfer directly to ACMIS: PRN velocity, paymentto-registration amount ratios, and reversal recency are analogous features exploited by the proposed agent. D. Insider Threat Detection The CERT Insider Threat dataset [5] established that insider threats exhibit characteristic precursors—elevated data access volume, off-hours activity, and privilege queries—detectable weeks before a damaging event. University attack surfaces expanded significantly during and after the COVID-19 pandemic, with documented surges in credential theft against educational institutions [6]. The proposed agent incorporates these precursor patterns as behavioural risk features. E. AI-Assisted Password Recovery Prior work on intelligent account recovery has focused on knowledge-based authentication and step-up verification [7]. The integration of NLP chatbots into the recovery flow— providing natural-language identity verification while simultaneously monitoring for abuse patterns—is a contribution of the present work. III. T HREAT M ODEL Table I enumerates the primary threat categories addressed by the proposed agent, classified by operational layer, example manifestation, and detection challenge. The threat model explicitly includes insider threats—actions by authenticated, authorised users who abuse legitimate access—which generate no authentication failures and do not violate access control lists. IV. S YSTEM A RCHITECTURE A. Overview The system follows a modular layered architecture (Fig. 1). A core AI security engine processes event streams from all ACMIS operational surfaces and produces risk scores. Sectorspecific modules define normal behaviour within each domain. A response orchestrator translates risk scores into automated actions. A dashboard exposes all signals to administrators in real time.
TABLE I ACMIS T HREAT M ODEL : C ATEGORIES AND D ETECTION C HALLENGES Layer
Example
Detection Challenge
Auth.
Brute-force (100 attempts / 2 min) Impossible travel (Kampala 10:00; Nairobi 10:20) Dormant account reactivation Privilege escalation
Distinguish typos from automated guessing Geolocation cross-session correlation
Auth.
Auth. Authoris. Financial Financial Behaviour Behaviour System Academic
Payment reversal after access grant Mass PRN generation (500 / min) Slow data exfiltration (10 records/hr) Bot mimicking human behaviour Logging / MFA disabled Simultaneous exam logins
Auth Events
Purely temporal; no active anomaly signal Each individual change may be authorised Temporal gap between payment and reversal Velocity anomaly; no per-PRN indicator Invisible at any single time point Speed signatures may be randomised No immediate attack; creates future risk Cross-session identity correlation
Financial Events
Behaviour Events
System Events
Core AI Security Engine (Anomaly Detection + Risk Scoring)
Academic Module
Financial Module
System Module
Response Orchestrator (Lock / Block / Alert / Escalate)
Admin Dashboard
Recovery Chatbot
Audit Log
Fig. 1. Modular architecture of the AI security agent. Blue: event inputs. Green: sector modules. Red: response orchestrator. Orange: output interfaces.
B. Core AI Security Engine The engine ingests structured ACMIS event streams via a message broker and scores each event through three parallel sub-models: 1) Sequence anomaly detector: An LSTM network trained on user action sequences. Sessions deviating from the user’s historical action distribution receive score sseq ∈ [0, 1]. 2) Statistical threshold monitor: A sliding-window velocity counter for high-frequency signals (login attempts, PRN generation, API calls), with class-specific thresholds θ learned from historical baselines. 3) Graph-based analyser: A graph neural network (GNN) modelling relationships between users, devices, IP ad-
dresses, and ACMIS entities. Unusual edges receive score sgraph ∈ [0, 1]. The composite risk score is: R = α sseq + β sthresh + γ sgraph ,
α+β+γ =1
(1)
D. Automated Response Framework Table II maps composite risk score R to a four-tier action set. Actions at all tiers are logged with full event provenance for forensic reconstruction.
where weights are tuned on validation data. R ∈ [0, 1] is mapped to a risk tier as described in Section V-D. C. Sector-Specific Modules Each module encodes domain knowledge about normal behaviour: Academic module—normal patterns for registration, exam login, and mark submission; Financial module—PRN velocity baselines, payment amount ranges, reversal recency windows; System module—expected API call rates, baseline configuration states (MFA enabled, logging active). D. Password Recovery Chatbot The chatbot is built on a retrieval-augmented NLP model fine-tuned for identity verification. The flow: (1) user initiates recovery and provides student/staff number; (2) OTP dispatched to registered contact; (3) chatbot verifies OTP and permits reset; (4) security engine simultaneously evaluates geolocation match, device novelty, and request velocity; (5) if anomalous, step-up challenge issued or session escalated. Mass reset events exceeding a configurable threshold trigger a system-level alert for potential credential takeover. V. D ETECTION AND R ESPONSE M ETHODOLOGY A. Brute-Force Attack Detection A sliding-window counter C(u, w) counts failed authentication events for user u within time window w: ( True if C(u, w) ≥ θbf Alert = (2) False otherwise Default: θbf = 10 failed attempts within w = 120 s. The source IP is simultaneously checked against a threat intelligence feed. B. Behavioural Baseline and Anomaly Scoring A baseline profile B(u) is built over 30 days per user, capturing login hours, session duration, action vocabulary, and geographic bounding box. A new session s receives: K
1 X ⊮[fk (s) ∈ / Bk (u)] sctx (s, u) = K
(3)
R
Low Medium
[0, 0.35) Log; update baseline [0.35, 0.65) Step-up MFA; rate-limit session [0.65, 0.85) Temp. lock account; block IP; pause transactions [0.85, 1] Emergency suspension; freeze financials; quarantine session
High Critical
Automated Actions
Notification None Dashboard (yellow) Push notification; incident report Immediate SMS/email; ticket raised
VI. E XPERIMENTS AND R ESULTS A. Dataset and Simulation In the absence of a publicly labelled ACMIS event log, a synthetic dataset was generated simulating normal ACMIS usage and injecting labelled attack scenarios. The dataset comprises 148,320 sessions across 2,400 simulated user accounts over 90 days, with a positive (attack) class prevalence of 8.3% (12,310 attack sessions across all nine threat categories). The 70/15/15% train/validation/test split was stratified by threat category and user role (student, lecturer, administrator, finance staff) to preserve class balance across splits. B. Baseline Comparisons The proposed agent is compared against: (1) Rule-based IDS (static thresholds as commonly deployed); (2) Isolation Forest [9] (standard unsupervised anomaly detector); (3) LSTM-only (sequence anomaly detection without graph and threshold components). C. Detection Performance
(4)
(5)
D. Response Latency
Impossible travel is flagged when:
where vmax = 900 km/h (air travel upper bound). C. Financial Transaction Monitoring PRN generation velocity: VPRN (u, w) = NPRN (u, w) / w
Tier
Table III presents per-category F1 scores on the held-out test set. The rule-based IDS scores near-zero on threats with no fixed threshold signature—impossible travel, slow exfiltration, insider threat, bot behaviour—confirming the inadequacy of rule-only approaches. The proposed agent achieves macroaverage F1 of 0.91, a 42-point improvement over the rulebased baseline and a 19-point improvement over LSTM-only.
k=1
dgeo /∆t > vmax
TABLE II F OUR -T IER AUTOMATED R ESPONSE F RAMEWORK
An alert fires when VPRN exceeds the 99th percentile of the user-class baseline. Payment reversals are flagged when reversal event er follows payment ep within ∆trev < 48 h and associated ACMIS services have already been unlocked.
Table IV shows mean automated response latency per risk tier. Critical-tier responses complete under 300 ms at the 95th percentile, within the sub-second threshold required for effective account lockdown before a brute-force attack succeeds.
TABLE III P ER -C ATEGORY D ETECTION P ERFORMANCE (T EST S ET, F1) Threat
Rules
Iso.F.
LSTM
Ours
Brute-force login Impossible travel Privilege escalation PRN mass generation Payment reversal fraud Slow data exfiltration Bot behaviour Dormant account abuse Insider threat
0.91 0.00 0.43 0.87 0.55 0.00 0.21 0.78 0.12
0.85 0.74 0.61 0.80 0.63 0.52 0.66 0.70 0.58
0.88 0.79 0.67 0.82 0.69 0.58 0.72 0.74 0.63
0.97 0.95 0.88 0.96 0.91 0.83 0.89 0.93 0.84
Macro-avg F1
0.49
0.68
0.72
0.91
TABLE IV M EAN AUTOMATED R ESPONSE L ATENCY BY R ISK T IER Tier Low Medium High Critical
Mean (ms)
95th pct (ms)
12 38 95 210
18 55 140 290
E. Chatbot Evaluation The password recovery chatbot was evaluated on 3,200 simulated recovery sessions: 74.2% normal recoveries and 25.8% injected mass reset attacks (defined as >15 reset requests within a 10-minute window across distinct accounts). Identity verification accuracy was 96.4% on legitimate sessions. Mass reset attack detection rate was 94.1%, with a false-positive rate of 3.8% on legitimate high-volume recovery periods (e.g., semester-start credential resets). Mean chatbot response time was 1.2 s per verification exchange. VII. D ISCUSSION A. Why Rule-Based Systems Are Insufficient Table III quantifies a well-known qualitative limitation: rulebased IDS achieves near-zero detection on threats invisible to threshold logic. Deploying only a rule-based system creates a false sense of security—administrators see brute-force alerts while a more damaging slow exfiltration proceeds undetected. B. The Modular Architecture Advantage The sector-specific module design addresses a key practical challenge: the definition of “normal” differs radically across institutional contexts. A financial module defines large transaction velocity as normal for a cashier but suspicious for a student account. Separating domain knowledge from the core detection engine allows retargeting to banking, healthcare, or government deployments by updating only the sector module. C. Deployment Considerations •
Warm-up period: Behavioural baseline requires 30 days of clean data; statistical threshold and graph components operate from day one.
Semester seasonality: Legitimate usage spikes at registration and examination periods; baselines must be seasonally adjusted to avoid false-positive surges. • Privacy: All event logs must be anonymised for storage; raw session data retained only during active incident investigation under institutional data governance policy. • Explainability: Administrator alerts include a humanreadable explanation of the behavioural features contributing to the risk score. •
D. Limitations The primary limitation is reliance on synthetic evaluation data. Real ACMIS event logs with confirmed attack labels are scarce due to institutional reluctance to share breach data. Future work will seek ethics-approved access to anonymised logs from a partner university. The LSTM model also requires retraining for new users; a meta-learning approach would reduce this cold-start latency. VIII. C ONCLUSION This paper presented an AI security agent for university ACMIS that addresses the fundamental limitation of rule-based intrusion detection: the inability to detect behaviourally subtle, temporally distributed, or contextually dependent threats. The proposed multi-layer agent—combining LSTM sequence modelling, statistical velocity monitoring, and graph-based relationship analysis—achieves macro-average F1 of 0.91 across nine threat categories, compared to 0.49 for a rule-based baseline. The four-tier automated response framework ensures highconfidence threats trigger protective action within sub-second latency, while low-confidence signals are logged without generating alert fatigue. The integrated NLP chatbot provides secure password recovery while monitoring the recovery flow for mass reset attacks. The modular architecture positions the system for extension beyond the university context: the core anomaly engine is sector-agnostic, with domain knowledge encapsulated in replaceable sector modules for banking, healthcare, and other institutional deployments. ACKNOWLEDGMENT The authors acknowledge the administration of Arapai Technologies International Limited, Uganda, for their domain expertise and operational insights into ACMIS security requirements, and commend their development of the AI security agent. R EFERENCES [1] I. Bongiovanni, “The least secure places in the universe? A systematic literature review on information security management in higher education,” Computers & Security, vol. 86, pp. 350–357, 2019. doi: 10.1016/j.cose.2019.06.012. [2] A. L. Buczak and E. Guven, “A survey of data mining and machine learning methods for cyber security intrusion detection,” IEEE Commun. Surveys Tutorials, vol. 18, no. 2, pp. 1153–1176, 2016. doi: 10.1109/COMST.2015.2494502.
[3] V. Chandola, A. Banerjee, and V. Kumar, “Anomaly detection: a survey,” ACM Computing Surveys, vol. 41, no. 3, article 15, pp. 1–58, 2009. doi: 10.1145/1541880.1541882. [4] M. Ahmed, A. N. Mahmood, and J. Hu, “A survey of network anomaly detection techniques,” J. Network and Computer Applications, vol. 60, pp. 19–31, 2016. doi: 10.1016/j.jnca.2015.11.016. [5] CERT Division, Software Engineering Institute, Carnegie Mellon Univ., “CERT insider threat dataset,” 2020. [Online]. Available: https:// resources.sei.cmu.edu/library/asset-view.cfm?assetid=508099 [6] H. S. Lallie et al., “Cyber security in the age of COVID-19: a timeline and analysis of cyber-crime and cyber-attacks during the pandemic,” Computers & Security, vol. 105, p. 102248, 2021. doi: 10.1016/j.cose.2021.102248. [7] J. Bonneau, C. Herley, P. C. van Oorschot, and F. Stajano, “The quest to replace passwords: a framework for comparative evaluation of web authentication schemes,” in Proc. 2012 IEEE Symp. Security and Privacy, San Francisco, CA, 2012, pp. 553–567. doi: 10.1109/SP.2012.44. [8] X. Yuan, P. He, Q. Zhu, and X. Li, “Adversarial examples: attacks and defenses for deep learning,” IEEE Trans. Neural Networks and Learning Systems, vol. 30, no. 9, pp. 2805–2824, 2019. doi: 10.1109/TNNLS.2018.2886017. [9] F. T. Liu, K. M. Ting, and Z.-H. Zhou, “Isolation forest,” in Proc. 8th IEEE Int. Conf. Data Mining (ICDM), 2008, pp. 413–422. doi: 10.1109/ICDM.2008.17.