Back to Paper
cs.CRcs.AIcs.LG
Local ID: 2603.28013v3
AI Summary: gemma4:e4b
Kill-Chain Canaries: Stage-Level Tracking of Prompt Injection Across Attack Surfaces and Model Safety Tiers
By Haochuan Kevin Wang, Zechen Zhang
Revision History Timeline
v13/30/2026
3/30/2026
No submitter comment provided.
v24/3/2026
4/3/2026
“10 pages, 8 figures. Benchmark code and run logs released”
v34/9/2026
4/9/2026
“10 pages, 8 figures. Benchmark code and run logs released”
★ Version indexed in Explorerv49/30/2026
9/30/2026
“12 pages, 6 figures, 6 tables. Code: https://github.com/KevinChunye/prompt_injection”
Comparing v3 vs v4
Green = Added • Red = Removed
Title Comparison
Kill-Chain Canaries: Stage-Level Tracking of Prompt Injection Across Attack Surfaces and ModelFive SafetyProduction TiersLLMs
Authors Comparison
No author changes.
v3 Comment
“10 pages, 8 figures. Benchmark code and run logs released”
v4 Comment
“12 pages, 6 figures, 6 tables. Code: https://github.com/KevinChunye/prompt_injection”
Abstract Word Diff
Multi-agent LLM systems are enteringnow productionread --documents, processingweb documents,pages managingand workflows,tool actingresults on behalf of users --users, yet their resilienceresistance to prompt injection is still evaluatedusually withreported aas singleone binary:number: did the attack succeed? This leaves architects without the diagnostic information needed to harden real pipelines. We introduce a kill-chain canary methodologymethod that tracksplants a cryptographicunique token throughin every injected payload and records the furthest of four stages (EXPOSEDit reaches (Exposed -> PERSISTEDPersisted -> RELAYEDRelayed -> EXECUTED)Executed), across 950 runs, five frontierproduction LLMs, six attack surfaces, and five defense conditions. TheExposure resultswas reframe100% promptamong injectionruns asthat acalled pipeline-architecturethe problem:tool; everythe modeloutcomes isdiffer fullydownstream. exposed,Claude yetHaiku outcomes4.5 divergeand downstreamClaude --Sonnet Claude4.5 blocksexecuted allnone injectionsof attheir memory-write164 (0/164text-surface ASR),attacks, and in the text relay the canary token never appeared in a memory write (0/40); GPT-4o-mini propagatesexecuted at53% 53%,of andits DeepSeekattacks. exhibitsFour 0%/100%findings acrossfollow. surfaces(1) fromA Claude writer kept the samecanary model.token Threeout findingsof mattershared formemory deployment:in (1)every write-noderelay placementrun we report; one cross-model pairing (Claude writer, GPT-4o-mini reader, n = 3) is consistent with this protecting the highest-leveragereader, safetyand decisionother --pairings routingwere writesnot throughtested. a(2) verifiedAs modelreaders, eliminatesthe propagation;Claude (2)models allexecuted four0/40 defensesraw failpre-seeded oninjections, atbut leastClaude oneHaiku surface4.5 dueexecuted to2/3 channelinjections mismatchrelayed alone,by noGPT-4o-mini; adversarialwhether adaptationrelayed required;injections (3)are invisibleharder whitefontto PDFrefuse payloadsthan matchraw orones exceedis visible-textan ASR,open meaningquestion. rendered-layer(3) screeningDeepSeek isChat insufficient.went Thesefrom dynamics0/24 applyon directlypre-seeded memory to production:8/8 institutionalon investorstool results, scenarios that also differ in task and financialpayload firmsformat; alreadywhite-text runPDF NLPpayloads, pipelinesinvisible overon earningsthe calls,rendered SECpage, filings,succeeded at least as often as visible ones. (4) pi_detector and analystwrite_filter reportsfailed --on channels they do not inspect, spotlighting failed on content it wraps, and write_filter blocked the document-ingestionPDF workflowsrelay nowbut migratingnot tothe LLMtext agents.relay, Code,a rundifference logs,we cannot explain. Code and toolingrun logs are publicly released.released: https://github.com/KevinChunye/prompt_injection