Software Security Monitors: Theory & Practice David Walker Princeton University (joint work with Lujo Bauer and Jay Ligatti) General-purpose Security Monitors • A security monitor (program monitor) is a process that runs in parallel with an untrusted application – monitors examine application actions • decide to allow/disallow application actions • may terminate an application, log application actions, etc. – monitors detect, prevent, and recover from erroneous or malicious behavior at run time – monitors generalize specific enforcement mechanisms such as access control lists, etc. July 2003 Software Security Monitors David Walker What is a security monitor? Monitors analyze & transform untrusted application actions: Input Stream … a3 a2 Monitor a2 a1 Output Stream … a4 a2 a2 a1 Application generates actions to be input into monitor Machine executes actions output by monitor July 2003 Software Security Monitors David Walker Possible Monitor Actions • Accept the action • Halt the application • Suppress (skip) the operation • Insert some computation • Also: replace results; raise exceptions July 2003 Software Security Monitors David Walker Formalizing security monitors • Security monitors => formal automata that transform a stream of program actions • Given: a set of possible program actions A • Monitors are deterministic state machines: (Q, q0, T) where Q = state set q0 = start state T = transition function July 2003 Software Security Monitors David Walker Operational Semantics Single step (determined by T): So (Sin, q) (Sin’, q’) Multi-step (reflexive, transitive closure of T): So (Sin, q) (Sin’, q) Output sequence is observable Input sequences are not observable July 2003 Software Security Monitors David Walker A Hierarchy of Security Monitors We classify monitors based on their transformational abilities (ie: based on T). Insert Suppress Truncation Suppression Insertion Edit July 2003 Software Security Monitors OK Halt David Walker An Example: E-Banana.com • Set of application actions: A={ } take(n), pay(n), browse, receipt // // // // take n bananas pay for n bananas browse for bananas commit • Edit Automaton: pay(n) browse browse start init take(n) pn tn take(n) pay(n) tpn receipt pay(n);take(n);receipt July 2003 Software Security Monitors David Walker Edit Automata Definition: (Q,q0,T) – where T = (t,e,i) – State transition function t • t : action x state state – Emission function e • e : action x state {+,-} – Insertion function i • i : action x state action sequence x state July 2003 Software Security Monitors David Walker Edit Automata • Operational Semantics a – (S, q) (S’, q’) if S=a;S’ and t(a,q)=q’ and e(a,q)= + – (S, q) (S’, q’) if S=a;S’ and t(a,q)=q’ and e(a,q)= Sins – (S, q) (S, q’) if S=a;S’ and i(a,q)=(Sins, q’) – (S, q) (empty, q) otherwise July 2003 Software Security Monitors (E-Accept) (E-Suppress) (E-Insert) (E-Halt) David Walker Security Policies • A program execution is a sequence of actions • A Security Property is a predicate over executions. • Example Properties: – P(S) iff bananas taken equal bananas paid for in S – Access control, resource bounds policies are policies • Non-properties: – Relations between different executions of a program – Information-flow policies July 2003 Software Security Monitors David Walker What does it mean to enforce a policy? • Principle of Soundness All observable outputs obey the policy sequences Sin . state q’ . sequence So So 1. (Sin, q0) (empty, q’) 2. P(So) • Principle of Transparency Semantics of executions that already obey policy must be preserved 3. P(Sin) (Sin So) July 2003 Software Security Monitors David Walker Some Useful Equivalences Remove/Insert unnecessary actions – fclose(f);fclose(f) fclose(f) • Replace a sequence with equivalent actions – socket(S);send(S,m) socketSend(S,m) • Permute independent actions – fopen(f);fopen(g) fopen(g);fopen(f) • Necessary properties: – reflexive, symmetic & transitive – S S’ P(S) P(S’) July 2003 Software Security Monitors David Walker E-Banana.com • Equivalence Rules: 1) (browse; S) S 2) (S1; take(n); pay(n); S2) (S1; pay(n); take(n); S2) July 2003 Software Security Monitors David Walker Conservative Enforcement Enforcer satisfies Soundness but not necessarily Transparency properties P . ( sequence S . P(S)) P can be conservatively enforced Conservative July 2003 Software Security Monitors David Walker Effective Enforcement Enforcer satisfies Soundness and Transparency provides some flexibility for the enforcer to edit the execution sequence guarantees the final results of running the application with the monitor are semantically equivalent to running the application without the monitor Effective Conservative July 2003 Software Security Monitors David Walker Precise Enforcement Definition Enforcer satisfies Soundness and Transparency Enforcer must output actions in lock-step with application Motivation In some scenarios, operations cannot be delayed without disrupting application semantics Precise Effective Conservative July 2003 Software Security Monitors David Walker What properties can be enforced? • The enforceable properties depend upon – the definition of enforcement (conservative, effective, precise) – the class of automaton (truncation, suppression, insertion, edit) – the space of possible input programs • if the monitor can assume certain “bad” executions do not occur, it can enforce more properties • static program analysis (type systems; proof-carrying code) can constrain program execution in ways useful to run-time monitors July 2003 Software Security Monitors David Walker Effective Enforcement • An E-Banana.com policy: browse*; ((take(n);pay(n) | pay(n);take(n)) ; receipt)* – Our edit automaton is an effective enforcer: • It satisfies Soundness • It satisfies Transparency • Proofs are by induction over the possible inputs – Less powerful automata (truncation, suppression and insertion) cannot enforce the E-Banana property • Proof by contradiction shows either Soundness or Transparency will be violated July 2003 Software Security Monitors David Walker A Simple Theorem • Theorem: Any decideable predicate P on executions is a property that can be effectively enforced by some edit automaton – Proof: construct a transactional edit automaton that suppresses and logs program actions when ¬P(S) and commits (outputs) when P(S), for every initial sequence of actions S in a program execution July 2003 Software Security Monitors David Walker Effectively Enforceable Properties Editing Properties Insertion Properties July 2003 Trunc. Prop. Software Security Monitors Suppression Properties David Walker Summary of theoretical results • We have developed the following rigorous methodology for reasoning about run-time security: 1. Define the computational framework using formal operational semantics 2. Define what it means to enforce a policy 3. Prove results about enforceable policies & mechanisms from definitions 1 & 2 July 2003 Software Security Monitors David Walker Future Work/Research Ideas • Proper definitions of enforcement for infinite execution sequences – Understanding edit automata on infinite sequences • Understand transactional policies & develop “transaction automata” – what can they enforce? • Incorporate more practical elements into the model – security environment; cryptographic secrets – replacement of results, exceptions and program state July 2003 Software Security Monitors David Walker Polymer, the Language • Polymer – A domain-specific language for programming security monitors (ie: edit automata) – Java + a couple of simple extensions: • atomic policy definitions encapsulating – a set of security-relevant actions – security state – decision procedure that produces security “suggestions” (halt, suppress action, insert action, etc) • compositional policy definitions involving – higher-order policy combinators July 2003 Software Security Monitors David Walker Securing Untrusted Applications untrusted code describes securityrelevant program points Java application policy interface instrumented application contains hooks to call monitor July 2003 separately compiled from policy Software Security Monitors David Walker Securing Untrusted Applications Java application policy interface policy implementation instrumented application combines application and policy July 2003 implements dynamic security policy secure application Software Security Monitors David Walker Atomic Polymer Policy class limitFiles extends Policy { private int openFiles = 0; private int maxOpen = 0; limitFiles(int max) { maxOpen = max; } new policy definition extends policy class private policy state .... } July 2003 policy constructor Software Security Monitors David Walker Atomic Polymer Policy Continued class limitFiles extends Policy { private int openFiles = ... private int maxOpen = ... public ActionPattern[] actions = new ActionPattern[] { <File fileOpen(String)>, <void fileClose(File)> }; .... set of policyrelevant methods } July 2003 Software Security Monitors David Walker Atomic Polymer Policy Continued class limitFiles extends Policy { private int openFiles = ... private int maxOpen = ... public ActionPattern[] actions = ... Suggestion before(Action a) { aswitch (a) { case fileOpen(String s) : if (++openFiles <= maxOpen) return Suggestion.OK(); else return Suggestion.Halt(); case fileClose(File f) : ... July 2003 Software Security Monitors policy behavior David Walker Atomic Polymer Policy Continued class limitFiles extends Policy { private int openFiles = ... private int maxOpen = ... public ActionPattern[] actions = ... Suggestion before(Action a) { aswitch (a) { case fileOpen(String s) : if (++openFiles <= maxOpen) return Suggestion.OK(); else return Suggestion.Halt(); case fileClose(File f) : ... July 2003 Software Security Monitors David Walker Atomic Polymer Policy Continued class limitFiles extends Policy { public ActionPattern[] actions = ... private int openFiles = ... private int maxOpen = ... Suggestion before(Action a) { aswitch (a) { case fileOpen(String s) : if (++openFiles <= maxOpen) return Suggestion.OK(); else return Suggestion.Halt(); case fileClose(File f) : ... July 2003 Software Security Monitors David Walker Complex Monitors • Combine atomic policies defined over a variety of different resources – eg: sample applet policy • file system access control • number of files opened • restricted network access – no network access after local file is read – communication with applet source only July 2003 Software Security Monitors David Walker Policy Combinators • Programmers may write parameterized policy combinators: – And, Or, Forall, Exists, Chinese wall,... AndPolicy: P1 s1 P2 s2 s July 2003 Software Security Monitors David Walker Policy Combinators class AndPolicy extends Policy { private Policy p1; private Policy p2; AndPolicy(Policy pol1, Policy pol2) { p1 = pol1; p2 = pol2; ... first-class } policies } July 2003 Software Security Monitors David Walker Policy Combinators class AndPolicy extends Policy { ... Suggestion before(Action a) { } } July 2003 Suggestion s1 = p1.before(a); Suggestion s2 = p2.before(a); if (s1.isOK() && s2.isOK()) return Suggestion.OK(); else ... using suggestions In reality, writing combinators is very tricky Software Security Monitors David Walker Summary of Language Design • Polymer facilitates the implementation of program monitors by 1. encapsulating all elements (relevant actions, state, decision procedure) of atomic policies in a single place 2. providing mechanisms to compose policies in a well-defined manner 3. coming equipped with a formal semantics – we’re working on it July 2003 Software Security Monitors David Walker Conclusions • Technology for securing extensible systems is in high demand – Software security monitors are one part of the solution • For more information, see – Edit Automata: Enforcement Mechanisms for Run-time Security Policies. IJIS 2003. – Types and effects for non-interfering program monitors. ISSS 2002 & LNCS 2609. – More Enforceable Security Policies. FCS 2002. – www.cs.princeton.edu/sip/projects/polymer/ July 2003 Software Security Monitors David Walker End July 2003 Software Security Monitors David Walker Realistic Monitors • Protect complex system interfaces – interfaces replicate functionality in many different places – method parameters communicate information in different forms – eg: Java file system interface • 9 different methods to open files • 4 different methods to close files • filename strings, file objects, self used to identify files July 2003 Software Security Monitors David Walker Abstract Action Definitions java.lang.io FileReader(String fileName); FileReader(File file); RandomAccessFile(...); ... FileReader.close(); RandomAccessFile.close(); ... July 2003 Software Security Monitors fileOpen(String n); fileClose(); David Walker Abstract Action Definitions class fileOpen extends ActionSig { boolean canMatch(Action a) { aswitch (a) { case FileReader(_) : return true; case RandomAccessFile () : return true; ... } } String parameter1(Action a) { .... } July 2003 Software Security Monitors David Walker Abstract Action Pattern Matching class limitFiles extends Policy { ... fileOpen.parameter1(a) Suggestion step(Action a) { aswitch (a) { case fileOpen(String s) : ... case fileClose() : } July 2003 } ... fileOpen.canMatch(a) Software Security Monitors David Walker Taxonomy of Precisely Enforceable Properties July 2003 Software Security Monitors David Walker Secure Application Untrusted application Host System (Java) Program Monitor Definition Polymer language extensions Java core July 2003 Software Security Monitors David Walker Policy Architecture: Simple Policies system interface Host System (Java) Simple Policy Def. Polymer language extensions Java core July 2003 Software Security Monitors David Walker Policy Architecture: Abstract Actions abstract system interface Host System (Java) concrete system interface July 2003 Abstract Action Def. Simple Policy Def. Polymer language extensions Java core Software Security Monitors David Walker Policy Architecture: Complex Policies Complex, System-specific Policy abstract system interface Host System (Java) concrete system interface July 2003 Abstract Action Def. Simple Policy Def. Policy Comb. Def. Polymer language extensions Java core Software Security Monitors David Walker Securing Extensible Systems • Many questions: – Our application requires property X. Can we enforce it precisely or will we have to get by with an approximation? – How do we write down our policy succinctly and unambiguously? – What specific mechanism will we need to enforce our policy? – How do we implement the mechanism? July 2003 Software Security Monitors David Walker Summary • A general framework for formal reasoning about security monitors – defined a hierarchy of security monitors – gave meaning to the word “enforceable” – developed rigorous proofs concerning enforceable properties • Polymer: A programming language for composing security monitors – techniques for modular monitor design & composition – formal semantics as an extension of FeatherWeight Java July 2003 Software Security Monitors David Walker