Tether: Verified Stateful Policy Engine for System-Level Monitoring of Agents
Organizations: Google DeepMind · University of Pennsylvania · Google · University of Wisconsin-Madison
Abstract
Autonomous AI agents are increasingly deployed for long-running tasks that execute arbitrary shell commands or programs. However, these actions can result in unauthorized network requests or persistent, malicious changes to the user's file system, processes, or privileges. Existing defense mechanisms rely on either LLM-based monitors, which lack provable safety guarantees, or stateless policy enforcement layers, which miss multi-step data flows and execution side-effects unmodeled by the policy language (e.g., network calls initiated inside a Python script). We introduce Tether, a framework for runtime enforcement of stateful security policies over the observed side-effects of an agent's actions. Tether executes each proposed action in an ephemeral, isolated sandbox, inferring abstract action semantics from observed system calls. Tether then enforces stateful policies over these inferred semantics, permitting actions to proceed only if they comply with the policy. Policies and action semantics are defined in an embedded DSL in Lean 4 capable of expressing diverse constraints, including information flow control (IFC), access control, and network security policies. Because actions can propagate labels from sources to destinations, a single rule can enforce fine-grained IFC policies across multi-step execution flows of arbitrary length. Expressing policies in Lean enables formal verification that they satisfy global security invariants (e.g., preventing network access after reading a sensitive file). We evaluate Tether on realistic coding-agent benchmarks across both benign and adversarial trajectories. With a single policy shared across all tasks, Tether blocks 100% of in-scope attacks while maintaining a benign false-positive rate under 5%.
Figures & tables
| Action | Updated State | Rule Name |
|---|---|---|
| E-Read | ||
| E-Recv | ||
| E-Write | ||
| E-Create | ||
| E-Delete | ||
| E-Copy |
| Lattices | \text{{\color[rgb]{0.06,0.46,1}deriving\_lattice}}\;L_{i}:= | ||
| \text{{\color[rgb]{0.06,0.46,1}label}}\;L:=L_{1}\times\dots\times L_{n} | |||
| Clearance | \text{{\color[rgb]{0.06,0.46,1}clearance}}\;\mathit{id}:L\;\text{{\color[rgb]{0.06,0.46,1}where}}\;\bar{\kappa} | ||
| Clearance rules | s\;\mathit{mode}\;[\text{{\color[rgb]{0.06,0.46,1}on}}\;k]\mapsto\ell | ||
| Policies | \text{{\color[rgb]{0.06,0.46,1}security\_policy}}\;\mathit{id}:L\;\text{{\color[rgb]{0.06,0.46,1}where}} | ||
| \hskip 8.50012pt\bar{D}\;\,\bar{c}\;\,\bar{R}\;\text{{\color[rgb]{0.06,0.46,1}default}}\;d_{0} |
| deriving_lattice C := Public < Internal < Sensitive < Secret |
| deriving_lattice I := Trusted < Untrusted |
| label Label := C I |
| clearance defaultClearance : Label where |
| "data/servers.csv" exact ( Sensitive , Trusted ) |
| "/id_rsa" suffixMatch ( Secret , Trusted ) |
| security_policy informationFlow : Label where |
| -- (a) Single fixed initial state ( start ) |
| security_check noSecretInit : Label where |
| start : clearance := initClearance |
| bad : (Taint.C Sensitive ) |
| -- (b) All safe initial states ( assume ) |
| security_check noSecretAccess : Label where |
| assume : Taint.C Sensitive |
| security_policy secretProtectionPolicy : Label where |
| set WorkspacePaths := [ ] |
| rule restrict_secret_read : on read( , ) |
| if [ , ].C Sensitive |
| then allow else deny "secret read" |
| rule restrict_secret_recv : on recvNet( ) |
| if [ , content].C Sensitive |
| Invariant | Description | Formal guarantee |
|---|---|---|
| No secret access | On any allowed run starting without Secret data, data labeled Secret never enters the agent’s context. | If , then and . |
| No tainted escalation | On any allowed run, every privilege escalation is denied after agent observes Sensitive , Secret , or Untrusted data. | If , then and along . |
| System integrity | On any allowed run, every action modifying, creating, or deleting a non-exempt protected system path is denied. | For every non-exempt protected path , for any action modifying , and for any attribute key . |
| No tainted protected edits | On any allowed run, once the agent observes Sensitive , Secret , or Untrusted data, every action modifying a protected path other than a cache path is denied. | For every protected non-cache path : if , then for any modifying ; if , then . |
| No exfiltration | On any allowed run, Secret data stays local; Sensitive data reaches only local hosts or app services. | ; . |
| Host blocklist | On any allowed run, every network send or file upload to a blocklisted host is denied, so no data ever reaches it. | For and : . |
| Benchmark | Monitor | Recall | FPR |
|---|---|---|---|
| Monitoring- BenchHonest | Tether | ( ) | ( ) |
| Auto Mode (default) | ( ) | ( ) | |
| Policy-aware Auto Mode | ( ) | ( ) | |
| LinuxArena | Tether | ( ) | ( ) |
| Auto Mode (default) | ( ) | ( ) | |
| Policy-aware Auto Mode | ( ) | ( ) |
| System | Observed Effects | Stateful Policies | Multi-hop IFC | Trace Proofs |
|---|---|---|---|---|
| Cedar ( Cutler et al., 2024 ; Disselkoen et al., 2024 ) | ✗ | ✗ | ✗ | e |
| OpenShell ( Watson, 2026 ) | a | ✗ | ✗ | e |
| Dogwood ( Brooker et al., 2026 ) | ✗ | c | d | ✗ |
| Strands Box + Dogwood ( Amazon Web Services, 2026 ) | b | c | d | ✗ |
| Tether | ✓ | ✓ | ✓ | ✓ |
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
| Abstract Action | Representative Syscalls / Sources | Emission Condition |
|---|---|---|
| read , pread64 , readv , mmap ( PROT_READ ), getdents64 , stat / fstatat , readlinkat , getxattr | Reads file contents ( ) or metadata ( ) into the agent’s context (either via stdout / stderr or via unconsumed process reads). | |
| openat ( O_CREAT ), creat , mkdir / mkdirat , mknod | Creates a path that did not exist before the step; holds the files previously read by the process whose data may influence ’s metadata. | |
| write , pwrite64 , writev , truncate , mmap ( MAP_SHARED ), chmod , chown , symlinkat | Mutates with no local source files ( ): for content writes/truncations, for chmod , for chown , and for symlinkat . | |
| unlink , unlinkat , rmdir , rename / renameat2 | Removes , or removes the old source path after copying to the new path during a rename . | |
| write , sendfile , splice , copy_file_range , mmap ( MAP_SHARED ), rename / renameat2 , linkat | Overwrites or links from a single source file (e.g., cp , mv , ln , or single-source shell redirection). | |
| write , pwrite64 , writev , copy_file_range , mmap ( MAP_SHARED ) | Overwrites from multiple sources ( ), or appends/modifies in place without truncation so ’s prior content survives ( ). |
AI Agent Security
Artificial intelligence agent security studies attacks and safeguards for tool-using, persistent, and collaborative systems across digital and physical settings.
See all 563 papers