Secure Autonomous Agents in SAP Systems: A Control-Property Review Evaluated on a Reconstructed SAP FI Authorization Model

{"Secure":[0],"Autonomous":[1,17],"Agents":[2,286],"in":[3,258,449,651,751,755,809,922],"SAP":[4,13,170,692,852,860],"Systems:":[5],"A":[6,179,421,557],"Control-Property":[7,157],"Review":[8,158],"Evaluated":[9],"on":[10,526,716,792,846],"a":[11,39,44,55,75,122,127,132,153,161,168,190,197,229,304,331,337,391,450,461,493,497,561,565,604,726,759,806,841,847,858],"Reconstructed":[12],"FI":[14,171,324,853],"Authorization":[15],"Model":[16],"AI":[18],"agents":[19],"are":[20,540,638,888],"being":[21],"given":[22],"authority":[23],"to":[24,33,52,302,477,517,520,569,593,666,730,766,906],"act":[25],"inside":[26],"enterprise":[27],"resource":[28],"planning":[29],"(ERP)":[30],"systems":[31],"—":[32,48,140,149,211,262,442,577,780,793,856],"post":[34],"the":[35,68,71,79,89,103,156,193,201,207,222,233,263,277,313,319,353,364,375,380,388,396,407,414,440,468,478,485,488,508,527,570,581,589,609,621,639,652,657,663,669,673,683,686,690,694,697,700,704,708,717,723,738,746,752,774,782,786,788,799,821,823,826,829,873,895,902],"credit":[36],"memo,":[37],"change":[38,568],"vendor's":[40],"bank":[41],"detail,":[42],"release":[43,877],"transport":[45,393],"into":[46],"production":[47],"rather":[49,910],"than":[50,221,232,241,616,773,911],"merely":[51],"suggest":[53],"that":[54,82,142,145,188,270,289,306,327,351,404,625,659,706],"person":[56],"do":[57],"so.":[58],"Almost":[59],"every":[60,265,349,475,649,925],"one":[61,136,542,551],"of":[62,99,163,369,418,427,434,580,612,762,802,850,897],"those":[63],"deployment":[64],"decisions":[65,298,403,412],"travels":[66],"with":[67,78,152,805],"same":[69,80,90,323],"reassurance:":[70],"agent":[72,458],"runs":[73,438,715],"under":[74,460],"named":[76,124,129,750],"user,":[77],"authorisations":[81],"user":[83],"already":[84],"had,":[85],"so":[86,359],"it":[87,134,213,227,250,437,523,576,617,757,899],"inherits":[88],"controls.":[91],"That":[92],"sentence":[93],"is":[94,108,113,181,192,204,239,275,328,399,492,620,642,644,654,740,745,771,790,794,817,820,838,879,927],"doing":[95],"an":[96,554,851],"enormous":[97],"amount":[98],"load-bearing":[100],"work":[101,530],"across":[102,299],"industry":[104],"right":[105],"now.":[106],"It":[107,112],"not":[109,246,487,496,504,681,736,857,880],"obviously":[110],"wrong.":[111],"also,":[114],"mostly,":[115],"untested.":[116],"This":[117,677,837],"study":[118,844],"tests":[119],"it.":[120,370,506,630],"For":[121],"single":[123,128,332],"control":[125,143,180,223,266,372,451,783,797],"against":[126,167,330,869],"property":[130,146,203],"at":[131,147,395],"time,":[133],"asks":[135],"deliberately":[137],"narrow":[138],"question":[139],"can":[141,373,424,626],"decide":[144],"all?":[148],"and":[150,160,196,295,431,500,522,547,549,578,601,636,648,662,672,703,713,776,798,828,833,876,882,886,901,935],"answers":[151],"diagnostic":[154],"instrument,":[155,822],"(CPR),":[159],"set":[162],"pre-registered":[164,695],"experiments":[165,314],"run":[166],"reconstructed":[169,691,848],"(Financial":[172],"Accounting)":[173],"authorization":[174,854],"surface.":[175],"The":[176,186,237,322,371,482,513,606,633,733,890],"central":[177],"idea.":[178],"neither":[182],"\\"strong\\"":[183],"nor":[184],"\\"weak.\\"":[185],"unit":[187],"carries":[189],"guarantee":[191,789],"control–property":[194,534],"pair,":[195],"pair":[198],"fails":[199],"whenever":[200,564],"protected":[202,571],"defined":[205,791],"outside":[206],"control's":[208,234],"observation":[209,769],"boundary":[210,770],"because":[212,226,436],"ranges":[214],"over":[215,406,785],"more":[216,611],"actions,":[217],"principals,":[218],"or":[219,225,917],"time":[220],"observes,":[224],"concerns":[228],"different":[230],"entity":[231],"predicate":[235],"names.":[236],"through-line":[238],"bigger":[240],"SAP:":[242],"autonomous":[243],"execution":[244],"does":[245,503,735],"create":[247],"these":[248],"gaps;":[249],"reveals":[251],"where":[252],"\\"local":[253],"compliance\\"":[254],"has":[255,865],"been":[256,867],"standing":[257],"for":[259,814],"\\"global":[260],"assurance\\"":[261],"assumption":[264],"environment":[267],"leans":[268],"on,":[269],"if":[271,668],"each":[272],"individual":[273],"transaction":[274],"compliant,":[276],"process":[278],"assembled":[279,341],"from":[280,342,656],"them":[281],"must":[282],"be":[283,518,627,907],"compliant":[284],"too.":[285],"don't":[287],"break":[288],"assumption;":[290],"they":[291],"move":[292],"fast":[293],"enough,":[294],"compose":[296],"enough":[297,300],"sessions,":[301],"make":[303],"gap":[305],"was":[307,515,524,543,599],"always":[308],"present":[309],"suddenly":[310],"visible.":[311],"What":[312,816],"find.":[315],"Three":[316],"measurements":[317],"anchor":[318],"paper:":[320],"Aggregation.":[321],"tolerance":[325],"group":[326],"complete":[329,842],"anomalous":[333],"document":[334],"cannot":[335,378],"discriminate":[336],"period-level":[338],"materiality":[339,357],"breach":[340],"individually":[343],"ordinary":[344],"documents.":[345],"Across":[346],"12,288":[347],"configurations,":[348],"setting":[350],"held":[352],"cumulative":[354],"outcome":[355],"below":[356],"did":[358],"by":[360,402,411],"refusing":[361],"legitimate":[362],"work;":[363,924],"best":[365],"retained":[366],"just":[367],"37.5%":[368],"throttle":[374],"total,":[376],"but":[377,685],"tell":[379],"two":[381,539],"workloads":[382],"apart.":[383],"Prevention":[384],"versus":[385],"detection.":[386],"On":[387],"transport-governance":[389],"surface,":[390],"harmful":[392,429],"identical":[394],"request":[397,415],"level":[398],"discriminated":[400],"only":[401,682,722],"range":[405],"transported":[408],"content,":[409],"never":[410],"reading":[413,467],"alone":[416],"(2,512":[417],"12,800":[419],"configurations).":[420],"governance":[422],"step":[423],"flag":[425],"40":[426,428],"releases":[430],"prevent":[432],"0":[433],"them,":[435],"after":[439],"import":[441],"perfect":[443],"detection,":[444],"zero":[445],"prevention,":[446],"both":[447],"green":[448],"matrix.":[452],"Attribution.":[453],"Under":[454],"principal":[455],"propagation":[456],"(an":[457],"acting":[459],"human's":[462],"just-in-time":[463],"identity),":[464],"no":[465],"procedure":[466],"recorded":[469,602],"audit":[470,559],"fields":[471],"beats":[472],"simply":[473],"assigning":[474],"action":[476],"busiest":[479],"actor,":[480],"unread.":[481],"record":[483],"names":[484],"person,":[486],"deciding":[489],"process;":[490],"this":[491,923],"schema":[494],"gap,":[495],"volume":[498],"one,":[499],"log":[501],"retention":[502],"repair":[505,778],"How":[507],"review":[509,885],"holds":[510],"itself":[511],"honest.":[512],"instrument":[514],"built":[516],"hard":[519],"please,":[521],"turned":[525],"authors'":[528],"own":[529,614],"first.":[531],"Of":[532],"eight":[533,582],"pairs":[535,583],"examined,":[536],"four":[537,579],"held,":[538],"contested,":[541,586],"withdrawn":[544],"as":[545,553,603],"misspecified":[546],"re-run,":[548],"exactly":[550],"survived":[552],"uncontested":[555],"mismatch.":[556],"specification-sensitivity":[558,824],"marks":[560],"verdict":[562],"\\"contested\\"":[563],"small,":[566],"defensible":[567],"property's":[572],"specification":[573],"would":[574],"flip":[575],"come":[584],"back":[585],"including":[587],"results":[588],"paper":[590,653,684,734,891],"had":[591],"hoped":[592],"claim.":[594],"One":[595],"headline":[596],"cost":[597,608],"result":[598],"retracted":[600],"loss.":[605],"method":[607],"authors":[610],"their":[613],"findings":[615],"confirmed,":[618],"which":[619],"most":[622],"trustworthy":[623],"thing":[624],"said":[628],"about":[629],"Reproducibility":[631],"discipline.":[632],"argument,":[634],"interpretation,":[635],"prose":[637],"authors';":[640],"what":[641,898],"mechanised":[643],"consistency.":[645],"Every":[646,711],"figure":[647,698],"number":[650],"generated":[655],"code":[658,946],"produced":[660],"it,":[661],"build":[664],"refuses":[665],"compile":[667],"written":[670],"claims":[671],"underlying":[674],"evidence":[675],"disagree.":[676],"deposit":[678],"therefore":[679],"contains":[680],"complete,":[687],"self-contained":[688],"testbed:":[689],"model,":[693],"experiments,":[696],"generators,":[699],"consistency-check":[701],"suite,":[702],"assembly":[705],"produces":[707],"manuscript":[709],"source.":[710],"experiment":[712],"check":[714],"Python":[718],"standard":[719],"library":[720],"alone;":[721],"figures":[724],"require":[725],"third-party":[727],"package.":[728],"Relationship":[729],"prior":[731],"work.":[732],"claim":[737],"principle":[739],"new.":[741],"Its":[742],"extent":[743],"form":[744],"classical":[747],"aggregation":[748],"problem,":[749],"multilevel-security":[753],"literature":[754],"1989;":[756],"specialises":[758],"published":[760],"characterisation":[761],"enforceable":[763],"security":[764],"policies":[765],"controls":[767],"whose":[768],"narrower":[772],"execution;":[775],"its":[777,893],"recommendation":[779],"give":[781],"state":[784],"horizon":[787],"history-based":[795],"access":[796],"mutable":[800],"attributes":[801],"usage":[803],"control,":[804],"real-world":[807],"instantiation":[808],"runtime":[810],"attribute-based":[811],"enforcement":[812],"products":[813],"SAP.":[815],"offered":[818],"here":[819],"audit,":[825],"measurements,":[827],"negative":[830],"results.":[831],"Scope":[832],"honesty":[834],"(please":[835],"read).":[836],"version":[839],"1.0,":[840],"independent":[843,883],"reported":[845],"model":[849],"surface":[855],"live":[859],"system.":[861],"No":[862,914],"modelled":[863],"mechanism":[864],"yet":[866],"verified":[868],"primary":[870],"vendor":[871],"documentation,":[872],"exact":[874],"product":[875],"scope":[878],"fixed,":[881],"practitioner":[884],"replication":[887],"planned.":[889],"states":[892],"scope,":[894],"limits":[896],"establishes,":[900],"external":[903],"validation":[904],"still":[905],"completed,":[908],"openly,":[909],"burying":[912],"them.":[913],"customer,":[915],"proprietary,":[916],"vendor-licensed":[918],"material":[919],"appears":[920],"anywhere":[921],"scenario":[926],"entirely":[928],"synthetic.":[929],"Authors.":[930],"Bindiya":[931],"Priyadarshini":[932],"(ORCID":[933,938],"0009-0005-7559-3896)":[934],"Martin":[936],"Pankraz":[937],"0009-0008-1567-4223).":[939],"Licensing.":[940],"Paper":[941],"text":[942],"CC":[943],"BY":[944],"4.0;":[945],"MIT.":[947]}

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-09-14
DOI
https://doi.org/10.5281/zenodo.22747364
Primary Topic
Robotic Process Automation Applications
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
article

Secure Autonomous Agents in SAP Systems: A Control-Property Review Evaluated on a Reconstructed SAP FI Authorization Model

Bindiya Priyadarshini, Martin Pankraz
Zenodo (CERN European Organization for Nuclear Research)
Robotic Process Automation Applications
article

Secure Autonomous Agents in SAP Systems: A Control-Property Review Evaluated on a Reconstructed SAP FI Authorization Model

Bindiya Priyadarshini, Martin Pankraz
article en

Abstract

Secure Autonomous Agents in SAP Systems: A Control-Property Review Evaluated on a Reconstructed SAP FI Authorization Model Autonomous AI agents are being given authority to act inside enterprise resource planning (ERP) systems — to post the credit memo, change a vendor's bank detail, release a transport into production — rather than merely to suggest that a person do so. Almost every one of those deployment decisions travels with the same reassurance: the agent runs under a named user, with the same authorisations that user already had, so it inherits the same controls. That sentence is doing an enormous amount of load-bearing work across the industry right now. It is not obviously wrong. It is also, mostly, untested. This study tests it. For a single named control against a single named property at a time, it asks one deliberately narrow question — can that control decide that property at all? — and answers with a diagnostic instrument, the Control-Property Review (CPR), and a set of pre-registered experiments run against a reconstructed SAP FI (Financial Accounting) authorization surface. The central idea. A control is neither "strong" nor "weak." The unit that carries a guarantee is the control–property pair, and a pair fails whenever the protected property is defined outside the control's observation boundary — because it ranges over more actions, principals, or time than the control observes, or because it concerns a different entity than the control's predicate names. The through-line is bigger than SAP: autonomous execution does not create these gaps; it reveals where "local compliance" has been standing in for "global assurance" — the assumption every control environment leans on, that if each individual transaction is compliant, the process assembled from them must be compliant too. Agents don't break that assumption; they move fast enough, and compose enough decisions across enough sessions, to make a gap that was always present suddenly visible. What the experiments find. Three measurements anchor the paper: Aggregation. The same FI tolerance group that is complete against a single anomalous document cannot discriminate a period-level materiality breach assembled from individually ordinary documents. Across 12,288 configurations, every setting that held the cumulative outcome below materiality did so by refusing legitimate work; the best retained just 37.5% of it. The control can throttle the total, but cannot tell the two workloads apart. Prevention versus detection. On the transport-governance surface, a harmful transport identical at the request level is discriminated only by decisions that range over the transported content, never by decisions reading the request alone (2,512 of 12,800 configurations). A governance step can flag 40 of 40 harmful releases and prevent 0 of them, because it runs after the import — perfect detection, zero prevention, both green in a control matrix. Attribution. Under principal propagation (an agent acting under a human's just-in-time identity), no procedure reading the recorded audit fields beats simply assigning every action to the busiest actor, unread. The record names the person, not the deciding process; this is a schema gap, not a volume one, and log retention does not repair it. How the review holds itself honest. The instrument was built to be hard to please, and it was turned on the authors' own work first. Of eight control–property pairs examined, four held, two are contested, one was withdrawn as misspecified and re-run, and exactly one survived as an uncontested mismatch. A specification-sensitivity audit marks a verdict "contested" whenever a small, defensible change to the protected property's specification would flip it — and four of the eight pairs come back contested, including results the paper had hoped to claim. One headline cost result was retracted and recorded as a loss. The method cost the authors more of their own findings than it confirmed, which is the most trustworthy thing that can be said about it. Reproducibility discipline. The argument, interpretation, and prose are the authors'; what is mechanised is consistency. Every figure and every number in the paper is generated from the code that produced it, and the build refuses to compile if the written claims and the underlying evidence disagree. This deposit therefore contains not only the paper but the complete, self-contained testbed: the reconstructed SAP model, the pre-registered experiments, the figure generators, the consistency-check suite, and the assembly that produces the manuscript source. Every experiment and check runs on the Python standard library alone; only the figures require a third-party package. Relationship to prior work. The paper does not claim the principle is new. Its extent form is the classical aggregation problem, named in the multilevel-security literature in 1989; it specialises a published characterisation of enforceable security policies to controls whose observation boundary is narrower than the execution; and its repair recommendation — give the control state over the horizon the guarantee is defined on — is history-based access control and the mutable attributes of usage control, with a real-world instantiation in runtime attribute-based enforcement products for SAP. What is offered here is the instrument, the specification-sensitivity audit, the measurements, and the negative results. Scope and honesty (please read). This is version 1.0, a complete independent study reported on a reconstructed model of an SAP FI authorization surface — not a live SAP system. No modelled mechanism has yet been verified against primary vendor documentation, the exact product and release scope is not fixed, and independent practitioner review and replication are planned. The paper states its scope, the limits of what it establishes, and the external validation still to be completed, openly, rather than burying them. No customer, proprietary, or vendor-licensed material appears anywhere in this work; every scenario is entirely synthetic. Authors. Bindiya Priyadarshini (ORCID 0009-0005-7559-3896) and Martin Pankraz (ORCID 0009-0008-1567-4223). Licensing. Paper text CC BY 4.0; code MIT.

Zenodo (CERN European Organization for Nuclear Research)
Peace, Justice and strong institutions
Openalex Percentile: Top 11%
Robotic Process Automation Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.