A target-specific authorization audit is introduced that labels context factors separately for each tool and argument target and holds the task, proposition, position, and policy fixed while changing only the proposition's source authority.
Abstract
LLM agents choose tools and arguments from context that mixes user requests, tool outputs, retrieved records, memory, and untrusted text. Evidence can be relevant without being authorized to determine a decision, so a correct action need not be grounded only in permitted evidence. We introduce a target-specific authorization audit that labels context factors separately for each tool and argument target. Its primary test holds the task, proposition, position, and policy fixed while changing only the proposition's source authority. We then test behavior when valid evidence is weakened and use context-subset interactions as a secondary localization diagnostic. Across 450 controlled next-action tasks and multiple open-weight LLM families, trusted and untrusted variants produce different actions in 5.4 percent of competing cases versus 1.7 percent of supporting cases. Under controlled degradation, unauthorized competition is retained in a full-correct, mixed-error, clean-correct pattern in 2.4 percent of comparisons, with a 95 percent confidence interval from 2.1 to 3.0 percent. These are controlled stress-set rates, not deployment prevalence. The models respond to textual source-authority cues, but this does not prevent untrusted evidence from influencing their actions.
This paper presents ToolGuardian, a policy-driven framework for securing agent-tool interactions through pre-admission vetting and task-aware runtime authorization, and compares ASP against heuristic and LLM-based policy realizations using identical inputs and output contracts.
Data cleaning without a trusted clean reference is challenging because unusual values may represent either genuine errors or valid observations. This paper studies how different agent capabilities affect reference-free data cleaning and proposes an evidence-grounded framework that combines structured context, profiling, LLM reasoning, executable checks, controlled evidence retrieval, source ranking, citation alignment, conservative repair, reversible scripts, and provenance logging. Seven configurations are evaluated across financial, clinical, and environmental-monitoring datasets using controlled synthetic corruption and original-data descriptive analysis, resulting in 126 completed runs. The evaluation includes two comparison baselines and a progressive LLM-based sequence that adds executable tools, evidence retrieval, evidence controls, and conservative repair. In the synthetic evaluation, the deterministic profiling baseline achieved the highest detection F1-score of 0.561. Among the LLM-based configurations, the full conservative configuration achieved the highest F1-score of 0.421, but no configuration performed best across all evaluation criteria. The source-ranked configurations achieved the lowest unsupported-rule rates, while decision-level citation alignment remained weak. The full conservative configuration produced no unsafe or unnecessary modifications, although these rates were already zero before the conservative policy was added, and it performed no direct repairs. Overall, the results show that additional capabilities introduce trade-offs among detection, repair, evidence grounding, conservative behaviour, reproducibility, and operational cost rather than producing consistent improvements. The study provides a structured framework and empirical methodology for evaluating these trade-offs in reference-free agentic data cleaning.
A correct answer can conceal why an agent succeeded. Once agents change their information state during evaluation, correctness no longer distinguishes intended reasoning from answer acquisition. Outcome evidence and exposure detection do not establish whether success depended on an acquired target; we call this missing evaluation object success provenance. AcquaBench audits it through matched CLEAN, GOLD, and SHAM value substitution on four standardized surfaces with joint qid-clustered analysis. CLEAN retains benchmark-authorized information. GOLD makes the correct target available. SHAM preserves source structure and exposure opportunity but substitutes a matched incorrect value. GOLD minus CLEAN measures the total score response to correct-target availability; GOLD minus SHAM tests whether that response tracks target correctness beyond matched source exposure. In D0, GOLD exceeds SHAM by 19.1 to 25.9 percentage points, showing that success follows the correct value. In D2, GOLD still exceeds SHAM under distributed sufficiency while coloc no longer transfers as a high-score marker, with AUROC 0.376 and 0.142. Behavioral dependence can thus persist beyond this probe's intended observation unit. In model comparison, a supported 5.0-point CLEAN score gap compresses to a raw GOLD difference of -0.6 points without establishing rank inversion. Agent benchmarks should report success together with whether the evaluated information state supported it.
A pipeline promoting an AI system publishes records claiming the thing evaluated is the thing deployed and that the evidence licensed the transition. We measure, from public material only, whether those records can express that claim and whether it holds where declared. First, a two-class documentation survey of 47 delivery platforms (20 CI/CD, 27 model-serving/agent) under one fixed three-label protocol, graded twice (second pass blind), every consulted page pinned by content hash and date. Across 188 double-graded cells we found no platform whose default record emits a content-addressed identity of the behavioral tuple (model version, instructions, tool definitions, retrieval and runtime configuration); the blind pass grades that column default on zero of 47. Immutable nominal versioning is meanwhile arriving as the agent platforms'default answer (16 of 27): version integers behind mutable pointers, a layer the artifact supply chain already found insufficient. Second, an instrument computes realized assurance depth from a pipeline's published exhaust alone and compares it with the declared depth. Applied to a frozen two-stratum frame of 30 public repositories graded twice from a hashed archive (second pass blind; cell-level agreement 23 and 19 of 30, both passes independently finding the same five full realizations), the sharpest result is a verifiability hole: seven of the 15 repositories chosen for adopting attestation tooling publish source-only releases, so the binding their workflows declare cannot be checked where declared. Where checkable it mostly checks out: five of seven measurable adopters realize the binding end to end; both shortfalls fall at identity binding. Together the results locate the field's records structurally short of justifiability, the one rung that can refuse a transition. The survey carries an expiry clock; we state what would falsify each finding.
Agent skills extend LLM agents with reusable procedures, tools, and domain-specific workflows, but their safety depends on resolving dependencies among interacting instructions. We introduce SkillLogic, a framework for analyzing logical relations in skill files and constructing executable tests from them. Our taxonomy covers eight relation types, including preconditions that gate valid actions, constraints that limit how allowed actions may be performed, and fallbacks that specify recovery behavior after failure. Using SkillLogic, we scan over 5000 public skills and find that 70% contain at least one logical relation. We then construct SLBench, an 86-case executable benchmark from high-confidence, high-impact, and locally testable relations. Evaluating Codex and Claude Code across six LLM backbones shows unsafe rates up to 70%, with violations leading to privacy leaks, unsafe configuration changes, and incomplete cleanup. The human audit attributes failures to both agent capability gaps and low-salience skill text. We further show that SLGuard, a lightweight inference-time scaffold, reduces violations by 63% on targeted cases. Our results establish logical-relation following as a distinct reliability challenge for skill-guided agents.
Xuan Chen, Chengpeng Wang, Lu Yan et al.· 0 citations
LLM agents routinely include privacy-sensitive data (PSD) in tool call arguments beyond what the invoked tools require, crossing trust boundaries to third-party services on every invocation. A controlled measurement on three production LLMs (GPT-4o, Claude 3.5 Sonnet, Llama-3.3-70B) shows that 81--88\% of tool calls include unnecessary PSD under default prompts; explicit privacy instructions still leave 36--76\% over-sharing. Existing defenses gate calls (allow/block) or label flows (information-flow control) but cannot \emph{rewrite} argument values, and PII detection tools miss implicit PSD like ``Memorial Sloan Kettering''(a hospital name that implies a diagnosis). We present \system{}, a middleware that intercepts tool calls and rewrites their arguments to the minimum data necessary for tool functionality, combining schema-aware necessity analysis with four operations: removal, generalization, substitution, and truncation. Live validation on 307 tool calls across the three LLMs above reduces privacy cost by 81.2--92.0\% at 100\% argument-level task validity (TOST equivalence $p{<}0.001$ at $\Delta{=}1.0$); on 25 unannotated Model Context Protocol (MCP) schemas, by 79.0\% with no \texttt{minimum\_necessary} metadata. An optional LLM content-necessity layer strips task-irrelevant PSD from otherwise-necessary free-text fields, raising live-LLM reduction to 85.1--95.6\% and author-schema reduction from 71.1\% to 90.9\%. Median latency is 1.77\,ms.