acceptodds
Under review as a conference paper at ICLR 2027

Surviving the Unseen: Predictive Defense for Novel Multi-Turn Multimodal Attacks

Abstract

The expansion of Multimodal Large Language Models (MLLMs) and their integration into autonomous agentic workflows has introduced a non-stationary attack surface. Empirical observations indicate that adversaries employ progressive, cross-modal perturbations that evade turn-specific guardrails by distributing malicious intent across longitudinal conversational trajectories. Static defense mechanisms, constrained by the Markov property, evaluate inputs in isolation and fail to detect cumulative structural poisoning. To handle this limitation, this paper formulates safety verification as a dynamic survival prediction and trajectory dynamics problem. The Triple-tier Anomaly Defense (TRIAD) framework is proposed as a predictive model that maps multimodal and multi-turn conversational flow as a continuous trajectory. The framework integrates structural anomaly detection to monitor covariance shifts, a Ledoit-Wolf regularized Mahalanobis distance to monitor covariance shifts in high-dimensional spaces, and topological trajectory acceleration to differentiate benign creative exploration from continuous malicious drift. These kinematic and geometric features are integrated into a time-varying Cox Proportional Hazards model via a Bayesian Hidden Markov Model (HMM) feedback loop. Theoretical analysis demonstrates that the TRIAD framework provides a mathematically bounded expected time-to-failure under adversarial perturbations, ensuring that malicious acceleration diverges positively. This framework provides a computationally efficient, interpretable, and predictive safeguard for real-time agentic AI systems, establishing a rigorous foundation for continuous safety alignment without relying on empirical retraining.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multimodal Large Language Models, Adversarial Attacks, AI Safety, Multi-Turn Jailbreaks, Agentic AI, Anomaly Detection, Survival Analysis

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_AnpKOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TRIAD: a theoretical framework for detecting adversarial attacks on LLMs in multi-turn, multi-modal environments. To detect attacks, the TRIAD framework: (1) defines a distance metric on the latent space of the model, (2) integrates it into a survival function, and (3) uses a hazard function to trigger an early stop if a threshold corresponding to safety-failure risk is exceeded. The paper presents two theoretical arguments that claim mathematical guarantees on the performance of the framework, which are accompanied by a proposed simulation protocol; however, no experimental results are reported.

Strengths:

  1. Addresses an important gap in LLM safety
  2. Claims mathematical guarantees on the performance of the framework
  3. Integrates behavioral signals, which could potentially help to distinguish safe from unsafe usage

Weaknesses:

  1. Deployment limitations are acknowledged but not sufficiently addressed. It is not established that the proposed mitigations resolve the limitations or preserve the claimed guarantees.
  2. Numerous grammatical errors make the manuscript difficult to follow at times.
  3. Mathematical contributions are not self-contained. Assumptions are not always clearly stated or motivated. Mathematical arguments are at times difficult to follow due to variables seemingly being re-used for different meanings.
Critical strengths

The main substantive strength is in addressing attacks that accumulate across multiple multimodal conversation turns. Prior work in this domain has demonstrated that intelligent attackers are capable of planning and executing their attack over multiple sequential turns. I believe that this form of attack will become more standardized, and therefore it is important to address methods of defense.

Critical weaknesses

The critical weaknesses are:

Line 134: "To statistically separate the mechanical characteristics of an automated attacker from the cognitive reflection of a normal user...": The behavioral distinction (mechanical characteristics of automated attacker from cognitive reflection of normal user) is insufficiently justified and does not address benign automation, human attackers, or automation that mimics human behavior.

Proof of Theorems 1 and 2: Regarding Theorem 1, the authors claim that a lower bound on one term of the distance formula applies as a lower bound on the entire squared-distance increment. It is both (a) not clear that this bound is correctly applied, and (b) not clear that the divergence test is applied correctly, and so it seems the proof does not adequately establish the claimed guarantee. Regarding Theorem 2, the authors appear to rely on a minimum speed assumption that is not justified. Theorem 2 appears to claim that if speed is above that minimum threshold, that this must imply strictly positive acceleration, but it is not clear that is true. For example, consider constant positive speed directly away from the reference. This would result in linearly increasing Mahalanobis distance and zero trajectory acceleration.

The gaps in the proof reduce my confidence in the paper's central theoretical contributions.

Questions for authors

Line 134: "To statistically separate the mechanical characteristics of an automated attacker from the cognitive reflection of a normal user...": What justifies separating malicious attackers from benign users based on automation-related behavior? Couldn't a malicious attacker be human (and therefore exhibit a human behavioral signature), or a benign user be automation (and therefore exhibit an automation behavioral signature)?

Re: Theorem 1: Would it be possible to elaborate on how the squared distance increment secures the given lower bound? I can see how the given bound derives from the last term in Eq. (11), but it seems like the cross term could be negative. I am also not clear how the divergence test applies here. Showing that a strictly positive exponential increment is added to the covariate is not the same as showing that terms fail to approach zero or that the sum approaches infinity.

Re: Theorem 2: How is this minimum speed assumption justified? Also, why does a positive ||V'(t)|| >= c > 0 for constant c imply that acceleration never converges to 0? Couldn't V'(t) be positive, constant, and directed away from \mu such that D_M increases linearly but trajectory acceleration is zero?

Re: the manifold heterogeneity section: Is it possible that an adversarial conversation could cause a change in the selected GMM cluster, reducing the measured distance and therefore artificially reducing acceleration under Eq. (7) to negative or zero despite the attack continuing? Could this possibly violate the claim in L258 that adversarial conversations exhibit "acceleration that strictly diverges in the positive direction"?

Reviewer_XEwtOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes TRIAD, a purely theoretical framework that models multi-turn multimodal MLLM safety as a survival analysis problem by cascading Isolation Forest anomaly detection, Ledoit-Wolf regularized Mahalanobis distance with trajectory acceleration, and a Cox Proportional Hazards model to predictively detect progressive cross-modal jailbreak attacks, without providing any empirical validation.

Critical strengths
  1. The manuscript is well-organized and readable.

  2. The experimental setup is described in a detailed and thorough manner.

Critical weaknesses
  1. The paper modifies the official ICLR 2027 LaTeX template. such as, widened margins.

  2. suspected AI-generated paper. The proofs largely restate definitions rather than derive non-trivial results.

  3. There are zero empirical results, no datasets, no baselines, no tables, no figures beyond a flow diagram.

Questions for authors

refer to weaknesses

Reviewer_RmTPOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes TRIAD, which casts safety monitoring of multi-turn multimodal conversations as survival analysis over a latent-space trajectory. Each turn's text and image embeddings are concatenated with a "behavioral covariates" vector (Eqs. 1–2); an Isolation Forest score (Eq. 4) gates a Ledoit-Wolf-regularized Mahalanobis distance to a fixed reference distribution (Eqs. 5–6) and its second finite difference, called "trajectory acceleration" (Eq. 7); an HMM belief over {Safe, Malicious} is updated; and a Cox proportional-hazards model (Eqs. 8–9) yields a hazard that, together with the sign of the acceleration, triggers early termination (Algorithm 1, Figure 1). Section 4 states two theorems: under continuously injected orthogonal perturbations the survival probability falls below any threshold in finitely many turns (Theorem 1), and adversarial drift produces strictly positive acceleration while benign exploration does not (Theorem 2). Section 5 lists deployment mitigations and describes a simulation protocol. The paper contains no experiments, datasets, baselines, or numerical results; the validation offered is the two theorems and the described-but-unexecuted protocol.

Critical strengths

The problem framing is reasonable and timely. Treating multi-turn jailbreaks as a time-to-event problem over an embedding trajectory, with a cheap unsupervised trigger gating a more expensive distance computation (Algorithm 1) and a Ledoit-Wolf estimator for the p ≫ n covariance (Eq. 5), is a sensible low-latency design that, if instantiated and evaluated, could become a usable stateful-guardrail baseline.

Critical weaknesses
  1. The two theorems that constitute the paper's entire validation do not hold as stated. Theorem 2 (p.6, lines 314–318) claims that under adversarial drift with ‖V′(t)‖ ≥ c > 0 the acceleration a_t "remains strictly positive" and "always exceeds a positive lower bound" (p.7, line 334). Constant-velocity drift V(t) = µ + t·u satisfies the proof's own assumption yet gives D_M(t) = t‖u‖_{Σ̂⁻¹}, linear in t, so the second difference in Eq. 7 is identically zero and Algorithm 1 line 10 ("h(t) > τ_hazard and a_t > 0") never alerts. The steps "∇D_M shows a monotonically increasing form" and "fails to turn negative due to the continuously added drift tension" (p.7, lines 327–333) are assertions, not derivations. In Theorem 1, the expansion in Eq. 11 contains the cross term 2[0,δ]ᵀΣ̂⁻¹_LW(V(t)−µ), which can be negative and exceed the quadratic term, so the claimed lower bound c‖δ‖² does not follow; and the "divergence test" argument (p.6, lines 308–312) uses only that each turn adds a positive amount to the cumulative hazard, which holds for any dialogue with h0(k) > 0, so a finite T_max exists with or without an attack and no explicit bound is derived. Since the abstract's central claims ("mathematically bounded expected time-to-failure", "malicious acceleration diverges positively") rest entirely on these results, their failure removes the paper's support.

  2. The method is not specified at a level that could be implemented, and no evaluation is provided. The reference (µ, Σ̂⁻¹_LW) is a fixed input to Algorithm 1 and the sample covariance (p.4, line 193) is over unspecified samples x_i — no dataset, encoder, or sample unit is named; B(t) is described only as "stochastic temporal variations, irregular cognitive pauses" (p.3, lines 141–143) and would not exist in the "autonomous agentic workflows" the abstract targets; the HMM (p.4, lines 204–208) has no transition or emission model and its belief (Algorithm 1 line 8) is never used by lines 9–10; β is "maximized during training" (Eq. 9) on labeled failure times whose source is never described, contradicting "without relying on empirical retraining" (abstract) and "unknown zero-day attacks" (p.2, line 57); and α and τ_hazard are never set. Despite the title "Practical Deployment and Experimental Validation Guidelines," Section 5 reports no experiment, and the closest cited prior work — DeepContext (Albrethsen et al., 2026), a stateful multi-turn detector with reported F1 — is used only to assert that guardrails are stateless (p.1, lines 47–48) and never compared against. Without a reproducible procedure or any measurement, the claimed "computationally efficient, interpretable, and predictive safeguard for real-time agentic AI systems" cannot be assessed.

Questions for authors
  1. For Theorem 2, how is the constant-velocity case V(t) = µ + t·u (which satisfies ‖V′(t)‖ ≥ c > 0 and yields a_t ≡ 0) excluded, and what is the mathematical content of "drift tension" and the claim that ∇D_M is monotonically increasing?
  2. For Theorem 1, how is the lower bound c‖δ‖² obtained when the cross term in Eq. 11 is negative, and what is the explicit expression for T_max as a function of ‖δ‖, β, and h0? Why does the divergence argument not apply equally to a benign dialogue?
  3. What data are used to estimate µ and Σ̂_LW, to fit the Isolation Forest, and to fit β in Eq. 9 (including how safety violations are labeled), and how is this compatible with detecting unseen attacks without retraining?
  4. What exactly is B(t) (features, dimension, collection), and how is it obtained when the "user" is another agent rather than a human typing?
  5. Have the authors run Algorithm 1 on any data, synthetic or real? If so, what detection/false-positive rates and latency were observed, and how do they compare to a per-turn guardrail and to DeepContext?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.