Disclosure Tactics Taxonomy

• Taxonomy of Disclosure Tactics is a structured classification system that organizes methods of controlled information release across varied domains.
• It categorizes approaches such as tool-based attacks, forum disclosures, financial reporting tactics, and compliance mechanisms with concrete examples from recent studies.
• The framework highlights both technical methodologies and socio-behavioral factors, demonstrating how ambiguity and staged revelations shape disclosure practices.

Searching arXiv for papers on disclosure taxonomies and related frameworks to ground the article. A taxonomy of disclosure tactics is a structured classification of the ways information is revealed, withheld, elicited, relayed, obfuscated, or made auditable. Recent research uses the term in markedly different senses: to classify intimate partner surveillance strategies, graded self-incrimination in cybercrime forums, illegitimate information flows in operating systems, voluntary reporting rules in financial markets, disclosure-avoidance mechanisms in official statistics, explanation requirements in automated decision-making, AI-use reporting in research, and discourse-functional tactics such as self-disclosure and validation in empathic language [(Tseng et al., 2020); (Hoheisel et al., 2 Mar 2026); (Hanspach et al., 2014); (Gietzmann et al., 2022); (Hawes et al., 10 Feb 2025); (Tsakalakis et al., 2022); (Ahmetoglu et al., 3 Apr 2026); (Gueorguieva et al., 9 Apr 2026)].

1. Domain structure and principal meanings

The literature does not treat disclosure as a unitary act. In some settings, disclosure is unauthorized revelation by an attacker; in others, it is a strategic choice by an informed party; elsewhere, it is a compliance artifact whose completeness, timing, audience, and granularity must be specified. This yields taxonomies that are categorical, dimensional, algebraic, or threshold-based, depending on the underlying problem.

Setting	Main categories or dimensions	Representative source
Intimate partner surveillance	Tool-based attacks that require physical access; tool-based attacks that do not require physical access; coercion and subterfuge; outsourced attacks	(Tseng et al., 2020)
Cybercrime forum disclosure	Benign; grey; crime-related; fallback “unclear”	(Hoheisel et al., 2 Mar 2026)
Voluntary financial disclosure	Candid; sparing	(Gietzmann et al., 2022)
Team and network disclosure	Full; partial; interior partial disclosure; ladder policies with posterior cutoffs	(Onuchic et al., 2023, Buhai, 28 Dec 2025)
Disclosure avoidance and explainability	Suppression, perturbation, rounding, thresholding, formal privacy; nine-dimensional explanation taxonomy	(Hawes et al., 10 Feb 2025, Tsakalakis et al., 2022)
Research and discourse reporting	AI disclosure elements and completeness criteria; ten empathic language tactics	(Ahmetoglu et al., 3 Apr 2026, Gueorguieva et al., 9 Apr 2026)

This heterogeneity suggests that “disclosure tactics” is best understood as a family resemblance term: the commonality lies in controlled information release, but the relevant units of analysis vary from forum posts and disclosure statements to posterior cutoffs, reporting cadences, and information-flow graphs.

2. Unauthorized revelation, concealment, and extraction

In intimate partner surveillance, the taxonomy is explicitly four-part. Tool-based attacks that require physical access include cellphone backup recovery tools, keyloggers and screen recorders, GPS tracking devices, and audiovisual recorders. Tool-based attacks that do not require physical access include leveraging shared phone plans, cloud or shared-account features such as iCloud and Google, network monitoring and web-traffic tracking, and reverse lookup directories and search. Coercion and subterfuge covers leveraging physical proximity for access, convincing the target to voluntarily provide access, and catfishing and social media manipulation. Outsourced attacks principally involve hiring private investigators (Tseng et al., 2020). The paper’s framing is significant because IPS is difficult to address precisely because of “the physical access and trust present in the relationship between the target and attacker” (Tseng et al., 2020).

In cybercrime forums, disclosure is neither fully criminal nor cleanly non-criminal. The three-level scheme distinguishes benign posts, grey posts in a legal or ethical grey area, and crime-related posts that overtly promote, sell, solicit, or exchange illegal goods, services, or criminal knowledge; a separate “unclear” label is used when context is insufficient. One quarter of initial posts include explicit crime-related content, more than one third of users disclose criminal activity at least once in their initial posts, over two-thirds post only benign or grey content, and grey initial posts are the largest single category at approximately $41\%$ (Hoheisel et al., 2 Mar 2026). The prominence of the grey category is important because the paper treats ambiguity itself as a tactic, functioning as a buffer zone and a source of plausible deniability (Hoheisel et al., 2 Mar 2026).

GenAI misuse work places disclosure-related abuse under the compromise of GenAI systems, especially attacks on data integrity and privacy. The relevant tactics are privacy compromise, data exfiltration, model extraction, social engineering via generative output, and prompt injection as an enabler of disclosure. The observed incidents indicate that privacy compromise and prompt injection are predominantly text-domain phenomena, while social engineering for self-disclosure extends into audio and multimodal settings; at the same time, the paper notes that actual in-the-wild forced disclosure attacks remain limited in public reporting (Marchal et al., 2024). In scientific disinformation, comparable tactics appear as concealment rather than extraction: concealing affiliations or political ties, creating a false persona, manipulating peer review, using vanity or predatory journals, and failing to disclose correction or retraction are all catalogued as methods of deceiving scholarly communication (McIntosh et al., 2023).

At the systems level, disclosure tactics can be formalized as undesired information flows. In the component-based operating-system taxonomy, legitimate flows include $p_i \Longleftrightarrow K$, $K \Longleftrightarrow H$, and $p_i \Longleftrightarrow K \Longleftrightarrow H$, whereas illegitimate flows include $p_i \Longleftrightarrow A$, $p_i \Longleftrightarrow p_j$, and $$p_i \Longleftrightarrow K \Longleftrightarrow H \Longleftrightarrow A$$. These are mapped to attack patterns such as insider attack, physical attacks, covert channels, policy flaws, and a novel covert physical channel that routes through the environment:

$$p_i \Longleftrightarrow K \Longleftrightarrow H \Longleftrightarrow E \Longleftrightarrow H \Longleftrightarrow K \Longleftrightarrow p_j$$

(Hanspach et al., 2014). In anonymous communication systems, the “family of statistical disclosure attacks” similarly treats disclosure as adversarial inference of communication patterns; the spectrum runs from OSDA and GSDA to weighted SDA, full-regression SDA, and LSDA, with LSDA outperforming the SDA family when the adversary has enough observations (Oya et al., 2019).

3. Strategic selective disclosure under incentives

Financial disclosure models treat disclosure tactics as equilibrium reporting rules. One paper defines two generic voluntary disclosure strategies between mandatory disclosure dates: a candid strategy, under which management unconditionally and immediately discloses every private news signal, and a sparing strategy, under which values are disclosed only if they exceed a dynamic threshold $\theta(t)$. The control variable is $T_t \in \{0,1\}$, with $p_i \Longleftrightarrow K$0 denoting candid and $p_i \Longleftrightarrow K$1 denoting sparing; mixed strategies are never optimal, optimal policies are of bang-bang type, and with constant pay-for-performance parameters at most one switching occurs (Gietzmann et al., 2022). The model characterizes the value process through

$p_i \Longleftrightarrow K$2

and gives a switching condition for candid-first strategies,

$p_i \Longleftrightarrow K$3

linking disclosure behavior to news intensity, pay-for-performance, and time to mandatory disclosure (Gietzmann et al., 2022).

A related investor-disclosure model yields a different taxonomy. Any investor who observes initial evidence $p_i \Longleftrightarrow K$4 always truthfully discloses it, but additional evidence $p_i \Longleftrightarrow K$5 is disclosed only if it is sufficiently extreme; moderate additional information is withheld. The paper characterizes this as a two-tailed reporting strategy and argues that it increases stock price volatility because investors profit from price movement in either direction rather than from the price level alone (Lu et al., 2024). The disclosure boundary is determined by thresholds $p_i \Longleftrightarrow K$6, with withholding when $p_i \Longleftrightarrow K$7 and disclosure outside that interval (Lu et al., 2024).

In team settings, disclosure is aggregated through a deliberation protocol. The taxonomy distinguishes full disclosure, partial disclosure, and interior partial-disclosure equilibrium. Full disclosure is unique only if the protocol allows every team member to unilaterally choose disclosure; partial disclosure exists if and only if not all team members can unilaterally choose disclosure; interior equilibrium exists if and only if no member can unilaterally force disclosure (Onuchic et al., 2023). Equilibrium strategies are threshold rules:

$p_i \Longleftrightarrow K$8

with possible mixing at equality, although the paper’s principal comparative statics are driven by unilateral, $p_i \Longleftrightarrow K$9-majority, consensus, and leader-follower protocols (Onuchic et al., 2023). The significant departure from classical unraveling is that partial disclosure can improve effort incentives under team-improving effort.

Dynamic network models add cadence choice to disclosure choice. Intermediaries choose whether to forward verifiable evidence and whether the public disclosure clock is on or off. When the clock is on, silence is informative and reputationally costly; when it is off, silence is mechanically uninformative. Along an expert–decision maker path, Markov perfect Bayesian equilibria are ladder policies with finitely many posterior cutoffs, and clock-off windows eliminate knife-edge mixing (Buhai, 28 Dec 2025). With sufficiently high reputational stakes and low discounting, dynamic incentives rule out persistent suppression and guarantee eventual transmission of all verifiable evidence along the path, even when bias reversals would block static unraveling (Buhai, 28 Dec 2025).

4. Disclosure avoidance, compliance, and structured reporting

In official statistics, disclosure tactics are protective mechanisms rather than reporting strategies. The framework for disclosure avoidance distinguishes inherent system features from implementation decisions and uses examples including suppression, perturbation or noise injection, rounding, thresholding, consistent post-processing, top- and bottom-coding, and formally private methods such as differential privacy with privacy-loss budget $K \Longleftrightarrow H$0 (Hawes et al., 10 Feb 2025). The paper stresses that candidate systems should be evaluated either with comparable parameter choices or across a range of parameter settings, because suppression thresholds, rounding granularity, noise magnitude, and $K \Longleftrightarrow H$1 are implementation decisions rather than intrinsic properties of the method (Hawes et al., 10 Feb 2025). It also states that no unified metric is proposed to equate every possible balance under the accuracy–availability–confidentiality tradeoff (Hawes et al., 10 Feb 2025).

Explainability research generalizes disclosure into a nine-dimensional taxonomy of explanations: source, perspective, autonomy, trigger, content, scope, explainability goal, intended recipient, and priority. Each explanation is classified across these dimensions, and the taxonomy is represented as a light OWL2 ontology with object properties such as hasSource, hasScope, and hasIntendedRecipient (Tsakalakis et al., 2022). This formulation makes disclosure requirements machine-readable and aligns explanation generation with regulatory and organizational compliance workflows.

In research reporting, DAISY operationalizes structured AI disclosure as a form-based tool. Its disclosure elements include AI tool name, tool version, location of AI use, purpose of AI use, extent or level of involvement, authorship or responsibility statement, and optional fields for human review and safeguards or input data shared. The six completeness criteria used in evaluation are tool name, tool version, locations of AI use, purposes of AI use, extent of involvement, and an explicit statement affirming human author responsibility (Ahmetoglu et al., 3 Apr 2026). DAISY-supported disclosures met more completeness criteria and did not reduce author comfort with the disclosure statements (Ahmetoglu et al., 3 Apr 2026). A plausible implication is that disclosure quality can be improved by converting vague policy norms into activity-based prompts.

5. Communicative, discourse-functional, and computational representations

Disclosure tactics are not confined to security, markets, or compliance. In empathic language, the taxonomy comprises ten tactics: emotional expression, empowerment, validation, information, paraphrasing, reappraisal, self-disclosure, advice, assistance, and questioning. These are mapped post hoc to three broader constructs—Experience Sharing, Perspective Taking, and Empathic Concern—and are defined at the level of phrase or discourse move rather than document-level report (Gueorguieva et al., 9 Apr 2026). Across two studies, LLM responses were found to be highly formulaic at a discourse functional level: a structured sequence of tactics matched between $K \Longleftrightarrow H$2 and $K \Longleftrightarrow H$3 of LLM responses, and between $K \Longleftrightarrow H$4 and $K \Longleftrightarrow H$5 in a held-out sample; when matched, the template covered $K \Longleftrightarrow H$6 to $K \Longleftrightarrow H$7 of the response (Gueorguieva et al., 9 Apr 2026). One of the paper’s regular-expression representations is

$K \Longleftrightarrow H$8

where $K \Longleftrightarrow H$9, $p_i \Longleftrightarrow K \Longleftrightarrow H$0, $p_i \Longleftrightarrow K \Longleftrightarrow H$1, $p_i \Longleftrightarrow K \Longleftrightarrow H$2, $p_i \Longleftrightarrow K \Longleftrightarrow H$3, and $p_i \Longleftrightarrow K \Longleftrightarrow H$4 denote tactic classes (Gueorguieva et al., 9 Apr 2026).

Operationalization methods vary sharply across domains. The IPS study uses mixed-methods qualitative analysis of online forums to surface tactics (Tseng et al., 2020). The cybercrime-forum study labels more than $p_i \Longleftrightarrow K \Longleftrightarrow H$5 million posts from nearly $p_i \Longleftrightarrow K \Longleftrightarrow H$6k users using an LLM pipeline with approximately $p_i \Longleftrightarrow K \Longleftrightarrow H$7 accuracy and $p_i \Longleftrightarrow K \Longleftrightarrow H$8 pairwise agreement against human-coded ground truth, then models transitions among disclosure states with Markov chains over $p_i \Longleftrightarrow K \Longleftrightarrow H$9 (Hoheisel et al., 2 Mar 2026). Statistical disclosure attacks in anonymous communication are represented as estimators; for LSDA the joint profile estimator is

$p_i \Longleftrightarrow A$0

(Oya et al., 2019). The operating-system taxonomy derives illegitimate flows by a recursive walk over a graph of entities (Hanspach et al., 2014). These differences are substantive: some taxonomies classify acts, some classify states, and some classify inference procedures.

6. Cross-domain themes and recurring debates

A recurring misconception is that disclosure is naturally binary. Multiple taxonomies reject that view. Cybercrime forums require at least a benign–grey–crime distinction (Hoheisel et al., 2 Mar 2026); team disclosure admits full, partial, and interior partial equilibria (Onuchic et al., 2023); financial reporting alternates between candid and thresholded sparing regimes (Gietzmann et al., 2022); investor disclosure separates always-disclosed initial evidence from selectively disclosed additional evidence (Lu et al., 2024). Taken together, these works suggest that ambiguity, thresholds, and staged revelation are not edge cases but central disclosure tactics.

A second misconception is that silence is always interpreted the same way. In dynamic networks, silence is informative only when the disclosure clock is on; when it is off, silence is mechanically uninformative (Buhai, 28 Dec 2025). In financial markets, the meaning of silence depends on the equilibrium disclosure regime and on the time to mandatory disclosure (Gietzmann et al., 2022). In investor-disclosure models, withholding moderate evidence can rationally amplify price reactions (Lu et al., 2024). These results make silence a state-dependent object rather than a uniform signal.

A third debate concerns whether disclosure is primarily technical or primarily social. IPS research emphasizes physical access, trust, coercion, and subterfuge (Tseng et al., 2020). GenAI misuse joins prompt injection and model extraction to social engineering via generated outputs (Marchal et al., 2024). Scientific disinformation combines concealment of affiliations, false persona, and publishing-process manipulation (McIntosh et al., 2023). DAISY identifies social, cognitive, and emotional barriers to reporting AI use, even when disclosure is normatively endorsed (Ahmetoglu et al., 3 Apr 2026). The consistent pattern is sociotechnical entanglement: technical affordances shape disclosure opportunities, but institutional and interpersonal incentives shape which tactics are actually used.

Finally, the literature disputes whether more structure necessarily deters disclosure. The disclosure-avoidance framework warns against “apples to oranges” comparisons between methods whose parameterizations differ (Hawes et al., 10 Feb 2025). Yet DAISY finds that structured, activity-based prompting increases completeness without reducing comfort (Ahmetoglu et al., 3 Apr 2026). This suggests that formalization can either constrain or enable disclosure, depending on whether it is used for punitive filtering, principled calibration, or recall support.