Skip to content

AI agent defenses: MCP-oriented decode-time provenance binding

Agentic AI Defense: Decode-Time Provenance Binding for MCP

arXiv· arXiv:2610.31974· Eric Yocam, Varghese Vaidyan, Paris Kalathas, Patrick Tague, Bruce DeBruhl, Yong WangDepartment of Computer Science, California Polytechnic State University; Beacom College of Computer and Cyber Sciences, Dakota State University; Information Networking Institute, Carnegie Mellon University; Department of Computer Science, University of IdahoIndexed by this site · 25 days ago· Originally publishedSelected DefensesSelectedAttention38

Quick read

Defenses

Based on the paper PDF (AI-generated); refer to the original

The authors propose decode-time provenance binding to defend against steganographic injection in MCP agents, reducing ASR from 100% (no defense) to 0% over 24 payloads.

Threat modelThe attacker can send the victim an email containing a steganographically signed logo, control external content and the behavior of all agents except the MCP layer and the verifier, and can read, replace, forge, or selectively present the binding; the attacker cannot obtain the signing key, cannot execute code within the MCP layer, and cannot edit the ledger. The victim system is a multi-agent MCP inbox summarization pipeline containing critical tools such as export_secrets and send_email.

Questions
Existing provenance defenses against indirect prompt injection apply trust labels in the text channel; the authors claim this can be bypassed by steganographic instructions in an email signature logo: the instructions enter the control flow only when a tool decodes them, which the authors call cross-modal trust laundering.
Methods
The authors propose decode-time provenance binding: the MCP layer mints an HMAC-signed Untrusted label when a tool returns, and independently maintains a provenance ledger and context taint for each critical call, recomputing at critical tool boundaries according to the ledger and blocking untrusted calls.
Experiments and results
On a real MCP testbed, with 24 adaptive payloads, no-defense ASR is 100%, the keyword scanner and regex parser each 62.5%, and this paper's mechanism 0%; a four-component ablation shows each component is necessary; trusted promotion raises legitimate task completion from 67% to 100%.
Limitations and future work
The authors state that the evaluation is a controlled testbed: 24 payloads, one cover logo, one small local model, lacking a production-grade email agent demonstration; the guarantee is limited to the execution context of full mediation, out-of-band paraphrase is out of scope; the defense protects actions rather than the summary text.
Experimental setup自建 MCP 测试台:24 条自适应载荷(3 种载体 × 8 种规避写法) · ASR、PSNR et al.
Benchmarks
自建 MCP 测试台:24 条自适应载荷(3 种载体 × 8 种规避写法)
Metrics
ASRPSNRSSIMlegitimate-task completionfalse blocksper-operation overhead (µs)
AI summaryand selection rationaleThe paper proposes a decode-time provenance-binding defense for MCP agents: at the moment the steganographic decoding tool runs, the MCP layer issues an HMAC-signed Untrusted label and independently maintains a provenance lineage of critical calls, blocking calls whose lineage is untrustworthy. The authors also characterize a cross-modal trust-washing attack: instructions are hidden via LSB, DCT, or metadata steganography in an email signature logo, evading scanners that inspect only text, and only become instructions when the tool decodes them, at which point they are washed into the trusted control flow of a multi-agent pipeline. On a real MCP testbed containing 3 carrier types and 24 adaptive payloads, the attack success rate was 100% for the undefended pipeline, 62.5% each for the keyword scanner and the regex parser, and 0% for this mechanism; the in-process taint baseline was likewise 0% within a single trust domain but failed across a compromised relay. The audited trusted-promotion path raised the legitimate task completion rate from 67% to 100%. The code and data are open-sourced on GitHub.

The paper proposes a decode-time provenance-binding defense for MCP agents: at the moment the steganographic decoding tool runs, the MCP layer issues an HMAC-signed Untrusted label and independently maintains a provenance lineage of critical calls, blocking calls whose lineage is untrustworthy. The authors also characterize a cross-modal trust-washing attack: instructions are hidden via LSB, DCT, or metadata steganography in an email signature logo, evading scanners that inspect only text, and only become instructions when the tool decodes them, at which point they are washed into the trusted control flow of a multi-agent pipeline. On a real MCP testbed containing 3 carrier types and 24 adaptive payloads, the attack success rate was 100% for the undefended pipeline, 62.5% each for the keyword scanner and the regex parser, and 0% for this mechanism; the in-process taint baseline was likewise 0% within a single trust domain but failed across a compromised relay. The audited trusted-promotion path raised the legitimate task completion rate from 67% to 100%. The code and data are open-sourced on GitHub.

Why it was selectedThe paper empirically demonstrates that text-channel provenance defenses can be bypassed by image steganographic payloads; its approach of binding trust labels at decode time and substituting containment for detection can be transferred to MCP agent protection.

Related events and news

1 events · 1 news mentions

News that mentions this paper

Adversa AI 汇总 2026 年 11 月 MCP 安全资源

Paper analysis

6 questions; open one to read the full answer

  1. What problem does this paper attempt to solve?

    The authors claim that provenance defenses in the text channel are bypassed by steganographic instructions in an email signature logo, and propose a defense that binds trust labels at decode time.

    Core problem: provenance defenses against indirect prompt injection all label the text channel; the authors claim this assumption can be broken by steganographic carriers—instructions hidden in images that enter the control flow only when a tool decodes them.

    • Problem scenario: An MCP agent reads external content and calls tools to act. An attacker steganographically hides instructions in an email signature logo (LSB, DCT, or metadata); the inbox-summary agent's decoding tool restores them into instructions, and in a multi-agent relay they are treated as ordinary intermediate results, which the authors call cross-modal trust laundering.
    • Shortcomings of existing methods: Content scanners look only at text and cannot see the payload in pixels; provenance tracking labels at text tool returns, whereas the dangerous transition happens at the decoding step—the authors claim both are bypassed by this carrier.
    • Core observation: Defenses should not detect whether an image hides poison (the authors call this an arms race the defender loses), but should assume any decoded output is untrusted and contain it, so that even a perfectly decoded payload cannot drive critical actions.
    • Paper proposes: decode-time provenance binding—the MCP layer mints an HMAC-signed Untrusted label at the instant the decoding tool runs, and independently maintains the full provenance lineage of critical calls; the verifier blocks calls whose lineage is untrusted. Also implements a real MCP testbed with three carriers and 24 adaptive payloads.
    • Threat model (Section III, V-D):
      • Attacker goal: Cause a critical tool (e.g., export_secrets) to execute when the true lineage contains an Untrusted decoding source and there is no audited promotion.
      • Attacker knowledge: Cannot access the victim's system prompt, model weights, keys, or execution environment.
      • Attacker capability: Can send the victim emails containing a steganographic signature logo; controls all external content and the behavior of every agent except the MCP layer and verifier; can read, replace, forge bindings, or selectively present handles.
      • Victim system: A multi-agent MCP inbox-summary pipeline, where a coordinator dispatches workers, at least one worker contains an image decoding step, and the critical tools export_secrets and send_email are reachable.
  2. What related research exists?

    Related work falls into four lines: indirect injection and benchmarks, steganographic injection, email/MCP attacks, and provenance and information-flow defenses; the authors state that this paper fills the gap at the intersection of the four.

    Indirect prompt injection and benchmarks

    • Greshake et al. (2023) demonstrated indirect prompt injection against real LLM-integrated applications, and subsequent work systematized it ([13], [14]).
    • Benchmarks such as InjecAgent, AgentDojo, and ToolEmu measure injection risk in tool-integrated agents; the paper notes that they established threats and evaluation methods, but all operate on the text channel.

    Steganographic injection against (multimodal) models

    • Spatial-domain (LSB), frequency-domain (DCT, DWT), and learned neural steganography can carry hundreds of characters visually imperceptibly ([18]–[20]).
    • Recent attack studies report 24–32% attack success rates against production-grade vision-language models, with cross-modal jailbreaks hiding payloads in images paired with benign prompts ([21]–[27]); the paper notes that these works feed a single constructed image directly to the model and do not consider email delivery, multi-agent pipelines, MCP tool calls, or provenance defenses.

    Email agents and MCP injection

    • Practitioners have repeatedly hijacked email-summary agents using text carriers such as zero-pixel HTML ([4]–[6]); NSA guidance documents MCP-specific risks such as poisoned tool outputs ([29]), and detectors and protocols exist for tool metadata poisoning ([30], [31]). The paper notes that these text carriers are naturally visible to text scanners, whereas this paper's image carriers are not.

    Provenance and information-flow defenses

    • Runtime policy enforcement points intercept MCP tool calls ([8], [9]); information-flow systems propagate taint between agent steps ([10], [11], [33], [34]); CaMeL separates trusted control flow from untrusted data flow ([35]); MELON provides provable constructions ([36]).
    • The paper names NeuroTaint [7] as the most directly related system and relays its self-described limitations: it does not consider multimodal attacks or semantic steganography, labels at text tool returns (rather than at decode transformation), labels are unsigned, and it audits completed traces rather than blocking at runtime.

    Baselines and positioning: The experimental baselines are a keyword/DLP content scanner, a regex intent parser, and an in-process data-flow taint tracker (which the authors explicitly describe as an unsigned design that trusts labels presented by the agent, not a SOTA provenance system). The authors state that this paper does not claim novelty in provenance or steganography per se, but in the decode-time combination of the two; Table I compares related systems along five axes, and the authors state that only this work simultaneously covers cross-modal carriers, decode-time labeling, signature-based tamper resistance, runtime enforcement, and multi-agent MCP.

  3. How does the paper address this problem?

    In a fully mediated MCP environment, an HMAC-signed label is minted at the instant a decoding tool returns; the layer maintains lineage itself, and critical calls are recomputed against the ledger and blocked.

    The overall idea is contain, do not detect: do not detect whether an image hides a payload, but assume any concealment-reversing tool's output is untrusted and contain it at the critical tool boundary. The mechanism is called decode-time provenance binding and consists of two algorithms, MintBinding (Algorithm 1) and EnforceCritical (Algorithm 2).

    Problem setting and two guarantees

    • Let D be the set of decoding tools (steganographic extraction, metadata decoding, Base64 decoding, OCR), and C the set of critical tools (sinks with side effects, such as secret export, outbound email, file write, payment).
    • The mechanism provides the conjunction of two guarantees: data-provenance security (critical actions cannot be authorized by dependencies on forged or relabeled data) and context-influence containment (once Untrusted instructions influence the execution context, critical actions cannot execute unless through audited promotion).

    Labeling and contagious labels

    • When tool t returns payload p from input bindings I, the MCP layer mints binding ℓ = (p, t, τ, π, δ, σ), where the dependency set δ is resolved by the layer from parameter handles it issued itself, not as claimed by the agent.
    • Trust labels are contagious: τ = Untrusted ⟺ t ∈ D ∨ ∃ i ∈ I: τi = Untrusted, i.e., a decoding tool's output or any untrusted input makes the result untrusted.
    • σ = HMAC_k(p∥t∥τ∥π∥δ) signs the entire record with the key k held by the layer; the agent receives only an opaque handle id(ℓ), not the key, and cannot mint or tamper with bindings.

    Ledger and critical-call enforcement

    • The layer maintains two structures itself: a binding store (mapping from handles to signed records) and a taint set per execution context (untrusted decoding sources that entered that context).
    • At a critical call, EnforceCritical does not consult the dependency set supplied by the agent, but rebuilds the transitive closure from its own ledger and recomputes every HMAC; if any binding fails verification, a dependency cannot be resolved, τ = Untrusted, or the context taint is non-empty (minus promoted sources), it Blocks.
    • Cross-agent messages are likewise mediated by the layer: the layer binds the message to its source and propagates taint to the target context. The authors explicitly state the scope boundary: when an agent paraphrases decoded instructions into free text and they leave out of band via an unmediated channel, this is not recorded.

    Security invariant and proof sketch

    • Invariant: in a fully mediated execution context, no critical tool call whose transitive lineage (data or context) contains an Untrusted decoding source executes, unless an audited trusted promotion clears that source.
    • Four assumptions: (i) the key k is confined to the MCP layer; (ii) every concealment-reversing tool is registered in D; (iii) the layer mediates every critical call; (iv) the layer records the full lineage of every call.
    • The proof sketch reduces security to the infeasibility of HMAC existential forgery (negligible probability within the key length) and to lineage-record integrity; the argument is information-flow style and does not inspect payload text, so the authors claim the guarantee is independent of wording and carrier.

    Legitimate tasks and deployment

    • trusted promotion: when a decoded product must legitimately drive an action (e.g., an invoice number from OCR), an explicit audit step authorized by a human or policy relabels that specific value as Trusted and records it in the context ledger.
    • Deployed as a drop-in component at the MCP server and tool-call boundary, with D and C being short allow/deny lists vetted at tool registration; across multiple hosts, shared keys or public-key signatures can replace HMAC.
  4. What experiments did the paper conduct?

    On a real MCP testbed with 24 payloads: no-defense ASR 100%, the two text baselines each 62.5%, this paper's mechanism 0%; ablation, replacement, overhead, and legitimate tasks are all evaluated.

    Experimental setup

    • Driven model: The agent is driven by a local Qwen/Qwen2.5-0.5B-Instruct; the authors state that safety decisions are made by a deterministic layer and do not depend on the model.
    • Testbed: Real FastMCP processes (stdio transport); the decoding tool mints bindings per Algorithm 1, and the layer maintains the ledger.
    • Attack library: 24 payloads = 3 carriers (LSB, DCT, metadata) × 8 evasion variants (plain, benign-worded, Base64, token-split, Cyrillic homoglyph, zero-width, Unicode-confusable, HTML-comment wrapping), designed specifically to defeat keyword and regex matching.
    • Attack flow: Each attack is delivered via an HTML-signed email; Worker1 decodes via a real MCP call, and the signed binding is relayed through peer hops and adjudicated by the verifier.
    • Compared configurations: No defense, keyword/DLP content scanner, regex intent parser, in-process data-flow taint tracker (IFC baseline), and this paper's mechanism—five in total.
    • Metrics: ASR (leakage rate over the 24 payloads); also PSNR/SSIM, legitimate task completion rate, and per-operation overhead; the paper does not report the number of repetitions.

    Main results

    Per Table II (24 payloads, lower is better):

    DefenseLeaksASR
    No defense24/24100%
    Content scanner (keyword/DLP)15/2462.5%
    Regex intent parser15/2462.5%
    Data-flow taint (IFC baseline)0/240%
    Decode-time provenance (this paper)0/240%
    • The authors state that the two text-channel baselines only stop payloads whose sensitive words survive Unicode normalization (plain, HTML-comment, and some confusable), and fail on benign-worded, Base64, token-split, homoglyph, and zero-width.
    • The authors explicitly state that this IFC baseline is an unsigned design that trusts labels presented by the agent, not a SOTA provenance system; the comparison is intended to isolate the two design properties this paper adds (cryptographic authentication and layer-maintained lineage).
    • The authors state that the difference between the taint baseline and this paper's mechanism is not ASR within a single trust domain but tamper resistance across untrusted relays, isolated by the ablation and cross-context experiments.

    Valid-label substitution and four-component ablation

    • Valid-label substitution (Table III): A compromised orchestrator holds a legitimate signed Trusted binding from a benign read_config and presents it as the argument to export_secrets, while omitting the Untrusted extract_stego output that actually drove the call; a naive verifier that trusts the binding presented by the agent lets it through (leak), while the layer-maintained check blocks it; in a clean context, legitimate critical calls are still allowed.
    • Four-component ablation (Table IV): The full mechanism blocks all three classes—honest attack, substitution attack, and cross-domain forged relay; removing decode marking or sink mediation leaks all three; removing HMAC still blocks the two in-domain classes but leaks the cross-domain forged relay; removing trusted lineage tracking still blocks forgery but leaks valid-label substitution. The authors conclude that cryptographic authentication and layer-maintained lineage are both necessary and not interchangeable.

    Cross-context propagation, relay tampering, and carrier distortion

    • Cross-context propagation (Table V): Agent A decodes an untrusted instruction and passes it to Agent B in another execution context via an intermediary message; the taint propagates and B's critical call is blocked; the non-intermediated bypass (A paraphrasing into free text) is marked as outside the model and not covered.
    • Relay tampering (Tables VII, XII): A genuine Untrusted label still verifies and blocks after 1/3/5/10 honest relay hops; a compromised worker rewriting the payload, forging a Trusted label, or stripping the binding is blocked due to HMAC mismatch; honest Trusted sources are allowed.
    • Carrier distortion (Table VIII): The metadata carrier is lossless (pixels unchanged, SSIM 1.000), LSB is near-lossless (PSNR ≈76 dB, SSIM ≈1.000), and DCT has PSNR ≈18 dB and SSIM ≈0.94. The authors emphasize that PSNR/SSIM are only distortion measures and do not constitute a claim of perceptual imperceptibility; formal subjective experiments are left for future work.

    Other ablations and analyses

    • Decoding-set completeness (Table XI): When D is correct, attacks are blocked and legitimate calls are allowed; when D lacks a decoder, attacks leak; when D is too broad, legitimate calls are falsely blocked.
    • Legitimate task utility (Table XIII): Of 3 legitimate critical tasks, 1 (an authorized payment using an OCR invoice number) is falsely blocked without promotion, giving a completion rate of 2/3 (67%) with 1 false block; with audited trusted promotion enabled, 3/3 (100%) with 0 false blocks, while same-origin attacks are still blocked.
    • Cryptographic overhead (Table IX): MintBinding ≈8.3 µs and EnforceCritical ≈8.7 µs, totaling ≈17 µs per critical call; relative to an ≈11 ms MCP decode round trip, this is about a 0.16% increase (measured over 2×10⁴ iterations after warm-up).
    • Full enforcement path (Table X): MintBinding stays stable at about 11 µs with closure depth; EnforceCritical grows linearly with closure depth—≈10 µs at depth 1, ≈0.68 ms at 100, and ≈6.6 ms at 1000—and is essentially flat with the number of concurrent tainted contexts (≈6 µs).
    • State transitions (Table VI): A clean logo has no critical call; without defense, export_secrets executes and the secret leaks; with defense, the payload is likewise decoded but the sink is blocked—the authors state that the decoded instruction is in context in both post-attack states, and the only difference is whether the critical call is permitted.
  5. What are some points that can be further explored?

    The authors list limitations including the controlled testbed, the fully mediated scope, and summary completeness; the experiments cover only one small local model and a self-built attack library.

    Limitations and future directions noted by the authors

    • Controlled testbed (Section IX): The evaluation is 24 payloads × 3 carriers × 8 evasions, one cover logo, one small local model, with keyword/regex scanners and an in-process taint tracker as baselines; a larger and more diverse payload library, multiple cover images and model families, modern learned or provenance-aware detector baselines, and a demonstration in an existing production email-agent workflow are all left to future work.
    • Scope limited to fully mediated contexts (Section V-C, IX): Effects that leave the mediated path (an agent paraphrasing decoded instructions into free text and passing them out of band) are not recorded; extending taint to in-model influence (tainting agent outputs when the input context is tainted) is left to future work.
    • Proof and formalization (Section IX): The proof is a sketch; a fully machine-checked proof and a formal treatment of lineage integrity are left to future work.
    • Protects actions only, not summaries (Section VIII, IX): A blocked injected instruction may still influence the natural-language summary the agent returns; quantifying the residual summary contamination rate is left to future work.
    • Design trade-offs (Section VIII): Containment cannot recover benign information that a decoded product may carry (e.g., a legitimate hidden watermark is likewise labeled Untrusted), requiring an explicitly audited promotion path; maintaining the decoding set D is an operational cost; keys are confined to the MCP layer, and cross-host use requires shared keys or signature verification; binding granularity is at the message/tool-result level and cannot express partial trust within a single decoded block.
    • Perceptual study (Section VI-H): PSNR/SSIM are only distortion metrics; a formal subjective perceptual experiment is left to future work.

    Experimental coverage

    • The model driving the agents is a single local Qwen/Qwen2.5-0.5B-Instruct; the authors state that safety decisions are made by the deterministic layer and do not depend on the model.
    • The attack library is 24 payloads (3 carriers × 8 evasion phrasings), using one cover logo, delivered via HTML signature email.
    • The compared configurations are five: no defense, keyword/DLP scanner, regex intent parser, in-process data-flow taint tracker, and this paper's mechanism; the authors explicitly state that the taint baseline is not a SOTA provenance system.
    • The legitimate-task evaluation is 3 critical tasks (Table XIII), the cross-context experiment is two agents (Table V), and relay integrity is tested up to 10 hops (Table XII).
    • The paper does not report the number of repetitions of the attack experiments; the overhead measurement is reported as 2×10⁴ iterations after warm-up (Section VI-I).
  6. Summarize the main content of the paper

    The authors propose a decode-time provenance binding defense against MCP steganographic injection: on 24 payloads, ASR drops from 100% to 0%, and legitimate task completion remains 100% via audited promotion.

    A defense paper against indirect prompt injection in MCP agents that also characterizes a class of cross-modal attack carriers.

    • Problem to solve: Existing provenance defenses apply trust labels at text tool returns; the authors claim steganographic instructions in email signature logos (three carriers: LSB, DCT, metadata) can bypass text scanners and enter the multi-agent control flow only when a tool decodes them, which the authors call cross-modal trust laundering.
    • Method highlights: decode-time provenance binding—the MCP layer mints an HMAC-signed Untrusted label at the instant the decoding tool runs, and the label is contagious; the layer independently maintains a binding ledger and a per-context taint set, and at critical calls rebuilds lineage from its own ledger and recomputes signature verification, without consulting the dependency set presented by the agent; the principle is containment rather than detection, so it is independent of payload wording and carrier.
    • Main results: On a real MCP testbed (FastMCP, agents driven by local Qwen2.5-0.5B-Instruct), ASR for 24 adaptive payloads: no defense 100%, keyword scanner and regex parser each 62.5%, this paper's mechanism 0% (Table II).
    • Key properties: Effective label replacement by a compromised coordinator is blocked by the layer-maintained lineage (Table III); a four-component ablation shows HMAC and layer-maintained lineage are necessary and not interchangeable (Table IV); labels remain intact across a 10-hop relay, and tampering, forgery, and stripping are all blocked (Table XII).
    • Utility and overhead: audited trusted promotion raises legitimate task completion from 2/3 (67%) to 3/3 (100%) while attacks remain blocked (Table XIII); cryptographic overhead is about 17 µs per critical call, about 0.16% of one MCP decode round trip (Table IX).
    • Authors' conclusions and boundaries: The authors claim that within a fully mediated MCP execution environment, containment at the decoding boundary plus layer-maintained lineage are the key properties that hold; the guarantee is limited to fully mediated contexts, out-of-band paraphrase is not covered, the defense protects actions rather than summary text, and a production-grade demonstration and machine-checked proof are left to future work.
Read originalarxiv.org