Beyond Bot Labels: Retrospective Consequence-Grounded Ranking of Attributed Cache Harm in E-Commerce Logs
Authors: Jisoo Kim, Youngwoo Seo, Seonghoon Jeong
Organizations: Department of Data Science, Sookmyung Women’s University, Seoul, Republic of Korea · Department of Statistics, Korea University, Seoul, Republic of Korea · Division of Artificial Intelligence Engineering, Sookmyung Women’s University, Seoul, Republic of Korea
Web caches reduce origin load and latency, but automated traffic can either warm them or evict objects shoppers reuse, lowering the cache hit ratio (CHR). Existing bot defenses generally infer automation from User-Agent (UA) strings or behavior rather than identify sessions that impose cross-session cache harm. We instead formulate the task as ranking sessions by eviction damage and propose consequence-grounded session ranking. A single chronological replay attributes each eviction-induced miss by a protected shopping session (a victim) to the session whose insertion last evicted the object. A behavioral model predicts the target from session behavior alone. Selected sessions bypass the cache but remain origin-served. In a controlled check of 539 non-victim sessions selected across harm levels, attribution order tracks victim-hit changes after removal (Spearman correlation 0.80). On a 12.7-million-request real trace, the top 1% of sessions carry 48.8% of attributed harm. Sessions without bot tags contribute 49.1% of this tier's harm. At the 20% request-volume target, the default conditional CHR difference reaches 93% of the UA baseline, which selects 47 times as many sessions. Six training seeds show little variation. This comparison conditions on the victims that remain, so we also report victim bypass. Local replay and retraining on a second store reproduce harm concentration and a top tier dominated by sessions without bot tags. These results establish attributed cross-session harm as an outcome-based target for retrospective cache audits and selective intervention design.
Figures & tables
Method
Analysis unit
Decision basis
Deng et al. (2008) [ 14 ]
client and object
repetition and lifetime
LRU-A (2003) [ 15 ]
client
recent own hit/miss history
DDCPC (2020) [ 16 ]
content-interest cluster
content-interest clustering
CacheShield (2012) [ 17 ]
object
observed request count
INCS (2024) [ 13 ]
flow
predicted synthetic bot class
Ours (2026)
session
predicted attributed harm
TABLE I: Comparison of traffic-aware cache defenses.
Fig. 1: Pipeline from raw traffic (❶) through preprocessing, attribution, and ranking (❷–❼) to bypass replay and CHR evaluation (❽–❾). Injected traffic enters stress tests only.
Fig. 2: Removal check for 539 sampled non-victim sessions. Attributed harm tracks victim-hit change (Spearman ρ=0.80 , 95% bootstrap interval [0.75,0.85] ). Axes are signed-log. The dashed line denotes identity.
Fig. 3: Attributed-harm concentration on EClog under 1% -capacity LRU. The top 1% and 10% carry 48.8% and 81% of harm (Gini 0.91 ). The diagonal denotes equality.
EClog
Zanbil
Observed phenomena
Singleton-URI share (cold tail)
54.5%
71.1%
Mean URI rarity, human → bot
3.3→5.6
3.4→6.8
Gini of per-session attributed harm
0.91
0.925
Top- 1% attributed-harm share
48.8%
67%
Top- 1% sessions without bot tag
41%
73%
TABLE III: Cross-store results and store-local bypass simulations.
Ranking
Sessions
Victim
Cond.
Victim
bypassed
bypass
Δ CHR (pp)
egress red.
Attributed-harm reference
3,637
0.27%
+1.20
5.8%
Harm ranking (ours)
3,551
0.18%
+1.04
5.2%
UA-based bypass
166,622
0.00%
+1.12
3.7%
LRU-A [ 15 ]
20,737
0.62%
+0.62
2.0%
INCS-style [ 13 ]
14,559
2.08%
+1.16
7.8%
TABLE IV: EClog bypass outcomes at 20% volume (LRU-A: threshold-selected 8.3% ). Victim-egress reduction compares post-bypass Vm origin bytes with no-bypass V .
Fig. 4: EClog conditional Δ CHR operating curves under LRU with a 1% cache. Breadth spans 1 – 20% bypassed volume. Bubble area denotes sessions. UA-based bypass has zero victim bypass by construction.
Method
ROC-AUC
Cond. Δ CHR (pp)
Sessions bypassed
Victim bypass
Contaminated LSTM
0.77±.04
+0.98±0.03
153 k
3.28%
Decontaminated LSTM
0.92±.02
+1.00±0.04
154 k
3.15%
Harm ranking (ours)
—
+1.02±0.02
3.5 k
0.18%
TABLE VII: Five-seed detector and six-seed harm-ranking outcomes (mean ± standard deviation).
AI-agent guardrails are memoryless: each message is judged in isolation, so an adversary who spreads a single attack across dozens of sessions slips past every session-bound detector because only the aggregate carries the payload. We make three contributions to cross-session threat detection. (1) Dataset. CSTM-Bench is 26 executable attack taxonomies classified by kill-chain stage and cross-session operation (accumulate, compose, launder, inject_on_reader), each bound to one of seven identity anchors that ground-truth "violation" as a policy predicate, plus matched Benign-pristine and Benign-hard confounders. Released on Hugging Face as intrinsec-ai/cstm-bench with two 54-scenario splits: dilution (compositional) and cross_session (12 isolation-invisible scenarios produced by a closed-loop rewriter that softens surface phrasing while preserving cross-session artefacts). (2) Measurement. Framing cross-session detection as an information bottleneck to a downstream correlator LLM, we find that a session-bound judge and a Full-Log Correlator concatenating every prompt into one long-context call both lose roughly half their attack recall moving from dilution to cross_session, well inside any frontier context window. Scope: 54 scenarios per shard, one correlator family (Anthropic Claude), no prompt optimisation; we release it to motivate larger, multi-provider datasets. (3) Algorithm and metric. A bounded-memory Coreset Memory Reader retaining highest-signal fragments at K=50 is the only reader whose recall survives both shards. Because ranker reshuffles break KV-cache prefix reuse, we promote CSR_prefix (ordered prefix stability, LLM-free) to a first-class metric and fuse it with detection into CSTM=0.7F1(CSDA@action,precision)+0.3CSR_prefix, benchmarking rankers on a single Pareto of recall versus serving stability.
Bot detectors deployed at scale treat traffic as binary: human or bot. This assumption breaks when AI agents browse the web through browser automation, a traffic class that is neither and that binary classifiers structurally cannot represent. We present a three-class detection framework distinguishing humans, bots, and AI agents, and show that the binary-vs-agent confusion is architectural: a binary human-vs-bot detector misroutes agent sessions because its label space lacks an agent class. On our controlled benchmark, an MLP binary classifier misclassifies 39.1% of real AI agents as human and a SAINT binary transformer misclassifies 34.5%; adding an explicit agent class yields per-class agent F1 = 1.000 in all 30 runs (3 model families × 10 seeds). To measure evasion resistance, we construct a five-level evasion ladder spanning passive observation, GAN-generated trajectories, and replay of real human cursor data (n=2299 evasion sessions). Across 10 seeds and 3 model families we observe zero agent misses in 22990 per-seed predictions. The discriminative signal is a browser-automation artifact, not evidence of agent reasoning: Playwright does not emit the raw pointer-move and wheel-delta streams a physical input device produces, and this absence signature survives trajectory manipulation. Exhaustive search over all feature subsets of size 1-5 (9401 GBMs) shows that two behavioral features (mouse_event_rate, teleport_click_ratio) give 100% observed agent recall at every evasion level with agent precision 0.994; five features lift macro-F1 to 0.991. The signal is redundantly encoded: removing teleport_click_ratio leaves agent detection at 100%. The single-feature regime is degenerate, flagging every agent only by collapsing the classifier to always predict "agent". Two features robustly isolate agents; five separate all three traffic classes at macro-F1 ≥0.99.
Vishisht Choudhary, Lukas Schmidt, Anne Zoë Kenntner +3
Technical University of Munich Munich, Germany · Kontext Munich, Germany
As LLM-based agents increasingly browse the web on users' behalf, a natural question arises: can websites passively identify which underlying model powers an agent? Doing so would represent a significant security risk, enabling targeted attacks tailored to known model vulnerabilities. Across 14 frontier LLMs and four web environments spanning information retrieval and shopping tasks, we show that an agent's actions and interaction timings, captured via a passive JavaScript tracker, are sufficient to identify the underlying model with up to 96% F1 within a fixed harness. Although fingerprints are harness-specific, with cross-harness transfer falling to near chance, pooling traces across harnesses recovers model identification at 0.74 to 0.81 Macro-F1. We further show that strong classifiers can be trained from roughly one third of the available traces, and that identification improves steadily as more events are observed. Injecting randomised timing delays between actions substantially degrades classifier performance, but does not provide robust protection: a classifier retrained on delayed traces largely recovers performance. We release our code and a labelled corpus of agent traces at https://github.com/KabakaWilliam/known_actions.
William Lugoloobi, Samuele Marro, Jabez Magomere +2
Oxford Internet Institute, University of Oxford · Department of Engineering Science, University of Oxford