DealingwithCodeThatIsOpaquetoSta7cAnalysis

BartonP.Miller†‡,JamesA.Kupsch†‡,ElisaHeymann†*,VamshiBasupalli†‡

NISTWorkshoponSoEwareMeasuresandMetricstoReduceSecurityVulnerabiliJes

Gaithersburg,MDJuly12,2016

†ComputerSciencesDepartment,UniversityofWisconsin‡DHSSoEwareAssuranceMarketplace(SWAMP)

*AutonomousUniversityofBarcelona

RecentExperience

•  ReviewedhighprofilevulnerabiliJes– Heartbleed(CVE-2014-010)– glibcDNSresolver(CVE-2015-7547)

•  Obtainedvulnerablesourcecode•  RanstaJccodeanalysistoolsoneach•  Toolsfailedtofindthebugs•  Bugwasopaquetothetools

2

HeartbleedAtit’sheart(sorry),it’sjustabufferoverflow…•  FailureoftheOpenSSLlibrarytovalidatetheheartbeatpacket

lengthfield(ascomparedtothesizeoftheactualmessage).•  HeartbeatpacketsarecontainedwithinTLSpackets.•  Theheartbeatprotocolissupposedtoechobackthedatasentinthe

requestwheretheamountisgivenbythepayloadlength.•  Sincethelengthfieldisnotchecked,memcpycanreadupto64KB

ofmemory.

memcpy(bp, pl, payload); Lengthfield.Suppliedbyanuntrustedsource.

Source.Bufferwiththeheartbeatrecord.Improperlyused.

DesJnaJon.Allocated,used,andfreed.OK.

3

TLSHeartbeatProtocol

4

type version len message

type len payload padding

Recordbuffer(allocatedbymalloc) ~16KB

TLSrecord

Heartbeatmessage

Illegallylargemessagelenallowsreading~64KB

Lenisthenumberofbytesofthepayload.Shouldbeconstrainedbythetotal

messagelength,header(3),andminimalpadding(16)

Lenisthetotalmessagelengthinbytes

Validread

5

Heartbleed

Addedlengthchecktoremediate:

if (1+2+payload+16 > s->s3->rrec.length) return 0 // silently discard

Andnoneofthecurrenttoolscouldfinetheproblem…why?

2556 unsigned char *p = &s->s3->rrec.data[0], *pl;

2563 n2s(p, payload); 2564 pl = p; 2565 2566 if (s->msg_callback) 2567 s->msg_callback(0, s->version, TLS1_RT_HEARTBEAT, 2568 &s->s3->rrec.data[0], s->s3->rrec.length, 2569 s, s->msg_callback_arg); 2570 2571 if (hbtype == TLS1_HB_REQUEST) { 2573 unsigned char *buffer, *bp; 2574 int r;

2580 buffer = OPENSSL_malloc(1+2+payload+padding); 2581 bp = buffer;

2584 *bp++ = TLS1_HB_RESPONSE; 2585 s2n(payload, bp); 2586 memcpy(bp, pl, payload);

Here’stheoffendingcode,slightlyredacted

6

2556 unsigned char *p = &s->s3->rrec.data[0], *pl;

2563 n2s(p, payload); 2564 pl = p; 2565 2566 if (s->msg_callback) 2567 s->msg_callback(0, s->version, TLS1_RT_HEARTBEAT, 2568 &s->s3->rrec.data[0], s->s3->rrec.length, 2569 s, s->msg_callback_arg); 2570 2571 if (hbtype == TLS1_HB_REQUEST) { 2573 unsigned char *buffer, *bp; 2574 int r;

2580 buffer = OPENSSL_malloc(1+2+payload+padding); 2581 bp = buffer;

2584 *bp++ = TLS1_HB_RESPONSE; 2585 s2n(payload, bp); 2586 memcpy(bp, pl, payload);

Here’stheoffendingcode,slightlyredacted

1.Findtheheartbeatpacketinthe(untrusted)userrequest

7

2556 unsigned char *p = &s->s3->rrec.data[0], *pl;

2563 n2s(p, payload); 2564 pl = p; 2565 2566 if (s->msg_callback) 2567 s->msg_callback(0, s->version, TLS1_RT_HEARTBEAT, 2568 &s->s3->rrec.data[0], s->s3->rrec.length, 2569 s, s->msg_callback_arg); 2570 2571 if (hbtype == TLS1_HB_REQUEST) { 2573 unsigned char *buffer, *bp; 2574 int r;

2580 buffer = OPENSSL_malloc(1+2+payload+padding); 2581 bp = buffer;

2584 *bp++ = TLS1_HB_RESPONSE; 2585 s2n(payload, bp); 2586 memcpy(bp, pl, payload);

Here’stheoffendingcode,slightlyredacted

2.Extractuser-statedpayloadlengthoftheheartbeatpacket

8

2556 unsigned char *p = &s->s3->rrec.data[0], *pl;

2563 n2s(p, payload); 2564 pl = p; 2565 2566 if (s->msg_callback) 2567 s->msg_callback(0, s->version, TLS1_RT_HEARTBEAT, 2568 &s->s3->rrec.data[0], s->s3->rrec.length, 2569 s, s->msg_callback_arg); 2570 2571 if (hbtype == TLS1_HB_REQUEST) { 2573 unsigned char *buffer, *bp; 2574 int r;

2580 buffer = OPENSSL_malloc(1+2+payload+padding); 2581 bp = buffer;

2584 *bp++ = TLS1_HB_RESPONSE; 2585 s2n(payload, bp); 2586 memcpy(bp, pl, payload);

Here’stheoffendingcode,slightlyredacted

3.plisanaliastotheheartbeatpayloadstartaddress.

9

2556 unsigned char *p = &s->s3->rrec.data[0], *pl;

2563 n2s(p, payload); 2564 pl = p; 2565 2566 if (s->msg_callback) 2567 s->msg_callback(0, s->version, TLS1_RT_HEARTBEAT, 2568 &s->s3->rrec.data[0], s->s3->rrec.length, 2569 s, s->msg_callback_arg); 2570 2571 if (hbtype == TLS1_HB_REQUEST) { 2573 unsigned char *buffer, *bp; 2574 int r;

2580 buffer = OPENSSL_malloc(1+2+payload+padding); 2581 bp = buffer;

2584 *bp++ = TLS1_HB_RESPONSE; 2585 s2n(payload, bp); 2586 memcpy(bp, pl, payload);

Here’stheoffendingcode,slightlyredacted

4.LengthofTLSpacketthatcontainsheartbeatpacket

10

2556 unsigned char *p = &s->s3->rrec.data[0], *pl;

2563 n2s(p, payload); 2564 pl = p; 2565 2566 if (s->msg_callback) 2567 s->msg_callback(0, s->version, TLS1_RT_HEARTBEAT, 2568 &s->s3->rrec.data[0], s->s3->rrec.length, 2569 s, s->msg_callback_arg); 2570 2571 if (hbtype == TLS1_HB_REQUEST) { 2573 unsigned char *buffer, *bp; 2574 int r;

2580 buffer = OPENSSL_malloc(1+2+payload+padding); 2581 bp = buffer;

2584 *bp++ = TLS1_HB_RESPONSE; 2585 s2n(payload, bp); 2586 memcpy(bp, pl, payload);

Here’stheoffendingcode,slightlyredacted

5.payloadlengthshouldbe≤TLSrecordlength-19

11

2556 unsigned char *p = &s->s3->rrec.data[0], *pl;

2563 n2s(p, payload); 2564 pl = p; 2565 2566 if (s->msg_callback) 2567 s->msg_callback(0, s->version, TLS1_RT_HEARTBEAT, 2568 &s->s3->rrec.data[0], s->s3->rrec.length, 2569 s, s->msg_callback_arg); 2570 2571 if (hbtype == TLS1_HB_REQUEST) { 2573 unsigned char *buffer, *bp; 2574 int r;

2580 buffer = OPENSSL_malloc(1+2+payload+padding); 2581 bp = buffer;

2584 *bp++ = TLS1_HB_RESPONSE; 2585 s2n(payload, bp); 2586 memcpy(bp, pl, payload);

Here’stheoffendingcode,slightlyredacted

6.allocateenoughmemoryforechopacket(accordinguserpayload)

12

2556 unsigned char *p = &s->s3->rrec.data[0], *pl;

2563 n2s(p, payload); 2564 pl = p; 2565 2566 if (s->msg_callback) 2567 s->msg_callback(0, s->version, TLS1_RT_HEARTBEAT, 2568 &s->s3->rrec.data[0], s->s3->rrec.length, 2569 s, s->msg_callback_arg); 2570 2571 if (hbtype == TLS1_HB_REQUEST) { 2573 unsigned char *buffer, *bp; 2574 int r;

2580 buffer = OPENSSL_malloc(1+2+payload+padding); 2581 bp = buffer;

2584 *bp++ = TLS1_HB_RESPONSE; 2585 s2n(payload, bp); 2586 memcpy(bp, pl, payload);

Here’stheoffendingcode,slightlyredacted

7.Copyheartbeatdatabasedonthelengththeyclaimed.Canalsograb

othernearbydata.

13

2556 unsigned char *p = &s->s3->rrec.data[0], *pl;

2563 n2s(p, payload); 2564 pl = p; 2565 2566 if (s->msg_callback) 2567 s->msg_callback(0, s->version, TLS1_RT_HEARTBEAT, 2568 &s->s3->rrec.data[0], s->s3->rrec.length, 2569 s, s->msg_callback_arg); 2570 2571 if (hbtype == TLS1_HB_REQUEST) { 2573 unsigned char *buffer, *bp; 2574 int r;

2580 buffer = OPENSSL_malloc(1+2+payload+padding); 2581 bp = buffer;

2584 *bp++ = TLS1_HB_RESPONSE; 2585 s2n(payload, bp); 2586 memcpy(bp, pl, payload);

Here’stheoffendingcode,slightlyredacted

Needtoactuallyknowthatpayloadlengthisnottrusted(tainted)data.

14

HeartbleedConceptually,thisisjustanexerciseintaintanalysis.WeneedtofollowingtheoriginalenclosingTLSpacketfromasocket,markingitastainted.Beforedisclosure:–  Notoolswetriedfoundthebug–  Notoolsweknowoffoundthebug

Coverity“fixed”theirtoolbynoJngthatextracJngtheintegerpayloadlengthfromanetworkbyte-orderusesabyte-swapinstrucJononaliqleendianmachine,andsuchaswapinstrucJonisrareenoughthatthisisasignthatthedatacomesfromthenetwork.

GrammaTechcoulddothetaintanalysisstarJngatsocketbuffers,butdidn’tdoitbecauseitwastooslowinpracJce.Whentheyturneditonfortherightsec.onofcode,itfoundtheproblem.

15

Difficul7esforSCATools

•  Legacylanguagesinherentfeatures– Rawmemoryaccess– Lackoftypesafety– Manualresourcemanagement– PointersandpointerarithmeJc

•  Codecomplexity–  IndirecJon– Largeprogramstate– Complexcontrolflow

16

WhySCAToolFailtoReport

•  NotdeducingaccuratesetofvaluesorproperJes(tainted,iniJalized,notnull,…)forvariables

•  NotdeducingcorrelaJonbetweenvariables•  UsingheurisJcstodeterminelikelyvaluesorproperJes

•  UncertainresultsnotreportedtoreducefalseposiJves

•  Confidencescoremaypointtoopaquecode,ifthereisareport

•  Fornon-reports,nowaytoconveyconfidence17

DynamicAnalysisTools

•  DynamicanalysisdidfindHeartbleed(singlefuzzedpacketcouldexposethevulnerability)

•  WedonotknowofanydynamicanalysistoolsthatfoundfoundglibcDNSvulnerability

•  DifficulJes:– GeneraJngcorrectbadinputsequence–  Inputdataspaceislarge–  Inputdatasequenceiscomplex

18

Goal:LessOpaqueCodeforSCA

•  Twoapproaches– Newcode:

•  Usemodernlanguagestopreventssomedefects–  D,Rust,modernC++

•  Use(more)analyzablesubsetoflanguage– MISRA–  CheckedC–  C++CoreGuidelines,GSL(guidelinesupportlibrary),andSCA

– Legacycode(andtoalesserextentnewcode):•  IdenJfypartsthatareopaque•  CurrentmetricsdonotidenJfyopaquecode

19

CommonMetrics

•  MetricTypes–  Simplecounts:

•  Lexicalelements:linesofcode,comments,…•  SyntacJcelements:parameters,types,operators,…•  PerfuncJon,file,orcodebase

–  Calculatedmetrics:•  Examples:CyclomaJc,Halsted•  PerfuncJon

–  RelaJonshipsbetweenfuncJons,classes,…•  Examples:Coupling,Cohesion,Connascence•  PerpairoffuncJons,classes,…

•  Inourexperience,thesemetricsdidnotcorrelatewithweaknessesorstaJcanalyzability

•  Focus:costtodevelop,maintain,test,enhance,…

20

Proposal:OpaquenessMetric

•  DeveloptoolsthatidenJfyprogramcomplexityintermsofopaquenesstoanalyzabilitybySCAtools–  SemanJccomplexityofcodethatreachesatool'sabilitytoreportduetoreachinglimitsoftheanalysisalgorithm's

•  Decidability•  ImplementaJon

–  Scoreregionsofthesourcecodewithanopaquenessscore– AlsoincluderaJonaleforpoorlyscoringregions

•  ProvideprescripJveadvicetotransformthecodetobelessopaquetoSCA(moreeasilyandcorrectlyanalyzable)

21

SCAToolProvidersPathForward

•  BestsemanJccodeanalysisisincommercialtools•  SCAtoolsalreadyhavemuchoftheinformaJon

– KnowwhereassumpJonsaremade– LocaJonofassumpJonsareaccurate– Shouldbeaccurateforusersofthetool

•  LimitaJons–  Inherentlynotintheirinterest,reporJnglimitaJonsisbadformarkeJng

– Specifictothetypesofproblemsthetoolfindsandthepowerofthetool

22

BroaderPathForward:DevelopTool

– StartwithexisJngopensourceanalysisframework•  ClangStaJcAnalyzer•  Gcc

– FundopensourcetoolbasedonframeworktoscorethesourcecodebasedonitsopaquenesstostaJcanalysis

– DevelopprescripJveguidanceontransformingsourcetomakecodelessopaque

23

Ques7ons

24