In Bouncy Castle for Java LTS before 2.73.13, the native one-shot CTR packet cipher did not check that the requested input length fitted the counter space the IV left. In CTR mode the IV and the block counter share one 16-byte block, so an IV of 13 to 15 bytes leaves a counter of only 1 to 3 bytes, addressing 256, 65536 or 16777216 blocks respectively. Given a longer input the counter wrapped and the keystream repeated from the start of the same packet, and the call then returned the full input length as though every byte had been correctly transformed. Two segments of the message were therefore encrypted under the same keystream, so their plaintexts can be recovered from the ciphertext alone, without the key, while the caller saw neither an exception nor a short length to indicate it. The streaming implementation validates at init and again while processing, and the portable AESCTRPacketCipher rejects such a request with "Counter in CTR/SIC mode out of range.", but the native one-shot path has a single entry point and performed no counter-range validation there. It now preflights the IV-derived counter range and rejects an over-long request before any output is written, so the operation is failure-atomic and never reports success for bytes it did not correctly transform. A counter of four bytes or more cannot be exhausted by a Java int length and is unaffected, as is a full 16-byte IV, where the counter range is the caller's responsibility. Bouncy Castle for Java (bcprov) is not affected, as it ships no native implementations.

In Bouncy Castle for Java before 1.86, validation of an MLS (RFC 9420) external commit's proposal list, org.bouncycastle.mls.protocol.Group.validateExternalCachedProposals, counted the proposals by type and bounded the removed leaf index but never established that the removed leaf had anything to do with the joiner. RFC 9420 sec. 12.2 permits at most one Remove proposal in an external commit, with which the joiner removes an old version of themselves, and requires that where one is present the LeafNode in the commit's path field meet the criteria it would have to meet in an Update for the removed leaf, in particular that its credential present identifiers acceptable for the removed participant. The ordinary proposal-list validator's self-remove rule is deliberately not applied on this path, because a resync commit legitimately removes a leaf the joiner owns, but nothing was put in its place. Any party holding the group's public GroupInfo, which is precisely what an external joiner is meant to be given, could therefore commit a Remove naming any member's LeafIndex and have every member apply it, evicting that member and taking over their slot in the ratchet tree. The credential check that should have prevented this existed only in the gRPC interop harness and so protected no other caller of the public Group.externalJoin and Group.handle API. An external commit carrying a Remove is now accepted only when the removed leaf's credential is identical to the one in the joiner's own new leaf, on both the sending and the receiving side.

In Bouncy Castle for Java before 1.86, BLS12381BasicScheme.keyValidate, and so BLSPublicKeyParameters and every BasicScheme, MessageAugmentation and ProofOfPossession verify and aggregateVerify that gate on it, accepted a public key built on a foreign ECCurve that merely shares BLS12-381's field characteristic. The prime-order subgroup check trusts a point's own curve to name its cofactor, since ECPoint.satisfiesOrder returns true outright when the curve's cofactor is one, so a point on a curve with a different equation and a cofactor forged to one passed keyValidate despite not being a G1 point at all. In BC's pairing implementation such a point contributes the identity in the target group, so an aggregate signature verified against a set of public keys including it is accepted even though it contains no signature for that key and message pair, admitting a phantom signer. keyValidate now first confirms that the point's curve carries exactly the canonical G1 field, equation, order and cofactor before any subgroup check. The issue is reachable only where an application constructs an ECPoint on an explicit, non-canonical curve and accepts it as an authority-bearing key; the standard 48-byte compressed-point decoder always supplies the canonical curve and was never affected.

In Bouncy Castle for Java before 1.86, HQC leaked secret-derived data through two side channels: its GF(2^8) arithmetic used lookup tables indexed by field elements, making the cache line touched a function of the operand, and its fixed-weight support sampler left its duplicate scan as soon as a collision was found and stored accepted positions at a secret index. Both run on secret inputs during encapsulation and decapsulation, and the sampler re-expands the secret key from its seed on every decapsulation, so an attacker able to observe cache behaviour or decapsulation timing can recover information about the HQC private key. The field arithmetic is now table-free and the sampler branch-free within a batch of candidates, with output and randomness consumption unchanged.

In Bouncy Castle for Java before 1.86, the Messaging Layer Security (MLS, RFC 9420) implementation did not bind an X.509 credential to a LeafNode's signaturekey. LeafNode.verify() checked a leaf's signature against the signaturekey carried in the leaf itself, while the credential's X.509 certificate chain was stored but never parsed or validated, so the end-entity certificate's public key was never required to match signaturekey as RFC 9420 sec. 5.3 requires. A party could therefore present another party's certificate as its credential while signing the leaf, and the enclosing KeyPackage, with an unrelated key, and be accepted under that other party's identity through KeyPackage.verify() and the Group leaf-validation path. In a deployment that admits external commits without an independent credential-admission check, an unauthenticated attacker could be admitted under a victim's X.509 identity, evict the victim (resynchronization compares whole credentials rather than signing keys), derive the current epoch, decrypt subsequent group messages, and send messages accepted as the victim. TreeKEM.LeafNode now requires the end-entity certificate's subject public key, in the cipher suite's signature encoding, to equal signaturekey for an X.509 credential and rejects the leaf otherwise, including an empty chain or a certificate whose key type does not match the cipher suite; certificate-chain and identity validation to a trust anchor remain the application's responsibility per RFC 9420 sec. 5.3.1. Deployments using only basic credentials are unaffected.

In Bouncy Castle for Java before 1.86, the high-level OpenPGP certificate API accepted a third-party certification or trust delegation from any component key of the issuing certificate, without requiring that component to have been granted the authority to certify. OpenPGPCertificate.getCertificationBy() and getDelegationBy() resolve a third-party signature by matching its issuer key identifier against every key of the third-party certificate, then verify the issuing component's binding chain and the signature itself; nothing checked that the issuing component carried the RFC 9580 sec. 5.2.3.29 certification key flag (CERTIFYOTHER) when the signature was created. A subkey bound only with SIGNDATA - the online signing subkey of exactly the offline-primary arrangement those key flags exist to express - could therefore issue a positive User ID certification over an attacker-controlled identity, or a full-trust depth-one direct-key delegation of introducer trust, and the API returned it as a valid signature chain attributed to the third-party certificate. An application treating getCertificationBy(...).isValid() or getDelegationBy(...) as an identity or trusted-introducer decision would attribute the attacker's assertion to the offline primary key. The same held for a legacy RSA subkey bound only for encryption, whose algorithm is nonetheless able to sign. This does not forge the primary key's signature or recover any private key; it promotes an already-compromised restricted subkey to the primary key's identity-issuing authority, defeating the containment the key-flag separation provides. A third-party certification or delegation is now attributed to the issuing certificate only when the component key that made it is the primary key, or is a subkey holding CERTIFYOTHER when the signature was created, so certification-capable subkeys continue to be accepted; primary keys are accepted whatever their key flags say, since a primary key is certification-capable by construction and certificates carrying no key flags subpacket at all are common. Third-party revocations are deliberately outside the rule, since declining to honour one would keep trust alive rather than withdraw it.

In Bouncy Castle for Java before 1.86, the high-level OpenPGP API accepted a data signature made by a signing subkey whose Subkey Binding signature carried no embedded Primary Key Binding (cross-certification) signature, in the case where that binding omits a Key Flags subpacket. RFC 9580 sec. 5.2.1.8 and sec. 10.1.3 require the embedded Primary Key Binding signature on any subkey that can issue signatures; it is the subkey's own statement that it belongs to the primary key it is bound under. OpenPGPCertificate resolved the subkey's key flags two different ways. isSigningKey() goes through getKeyFlags() and getApplyingSubpacket(), which falls back to the primary key's direct-key or primary User ID self-signature when the binding signature omits the subpacket, so the subkey inherited the primary's SIGNDATA and counted as signing-capable; verifyEmbeddedPrimaryKeyBinding(), which enforces the requirement, reads the binding signature's own hashed subpackets, found no SIGNDATA there, and returned early as a non-signing key without ever demanding the back signature. The same subkey was therefore signing-capable - so its signatures were attributed to the certificate and OpenPGPSignature.OpenPGPDocumentSignature.isValid() returned true - while being exempt from cross-certification, where GnuPG refuses the identical certificate and message. An attacker needs only the victim's public signing subkey, which is public material: they bind it to their own primary key with a Subkey Binding signature they are able to make, carrying no Key Flags and no embedded Primary Key Binding signature, which they cannot make without the subkey's private key, and a relying party verifying one of the victim's genuinely signed messages against that certificate is told the signature is valid and given the attacker's certificate as its issuer. Because a certificate's User IDs are self-asserted, a verifier that pins on the subkey's fingerprint or key ID while taking the identity from the enclosing certificate reports a real signature under an attacker-chosen identity. This is misattribution of a genuine signature rather than forgery of a new one: no private key is recovered, and the signature must be one the grafted subkey actually made. The low-level PGPSignature / PGPPublicKeyRing API performs no binding checks by design and is unaffected. Key Flags are a statement about the key the carrying signature refers to (RFC 9580 sec. 5.2.3.29), so a subkey no longer inherits them from the certificate-wide signatures of the primary key: a Subkey Binding signature that omits the subpacket now leaves the subkey with no capabilities rather than the primary's, which makes the flags the cross-certification check consults the same flags every other decision consults. Preferences and the other subpackets a direct-key signature carries are inherited as before, and the primary key itself, whose flags legitimately come from its own direct-key or User ID self-signature, is unaffected.

In Bouncy Castle for Java LTS before 2.73.13, the one-shot native packet ciphers for AES-CBC, CCM, CFB, CTR, GCM and GCM-SIV released the caller's key, IV and additional authenticated data arrays with JNI's ReleaseByteArrayElements in mode 0, which commits the native copy back into the Java array. Those arrays are read-only to the native code, and on a JVM that returns a copy rather than a pin the copy still holds the input bytes as they were read. The output buffer is taken through a separate critical region and committed first, so where an application passed the same Java array as both an input and the destination - encrypting in place over KeyParameter.getKey(), for example - the later mode-0 release of the key wrote the unchanged key bytes over the ciphertext that had just been produced. The call still returned the correct output length, so an application encrypting in place over its own key array was handed the raw AES key where it expected ciphertext, with nothing in the API to indicate it, and would transmit or store the key in place of the message. The read-only input arrays are now released with JNIABORT, freeing the native copy without copying it back, and mode 0 is reserved for arrays the native code wrote. The pure-Java packet ciphers and the streaming native modes are not affected. Bouncy Castle for Java (bcprov) is not affected, as it ships no native implementations.

In Bouncy Castle for Java before 1.86, the streaming CMS AuthenticatedData parser accepted a message whose digestAlgorithm and authAttrs fields disagreed about whether authenticated attributes were present. RFC 5652 sec. 9.1 pairs the two, requiring that authAttrs be present whenever digestAlgorithm is, and sec. 9.2 makes the MAC cover the DER encoding of authAttrs when they are present and the eContent OCTET STRING directly when they are not. CMSAuthenticatedDataParser has to choose between those two in its constructor, before it can reach authAttrs, which comes later in the SEQUENCE, so it chose on digestAlgorithm alone: for a message with digestAlgorithm absent but authAttrs present it verified the content MAC and then returned the attributes through getAuthAttrs() as though they had been authenticated, when the MAC had never covered them. An attacker able to modify a message in transit could insert an authenticated attribute, such as an RFC 2634 ESSSecurityLabel, into an otherwise valid message while holding neither the key-encryption key nor the content-MAC key, and an application taking an authorization, routing or labelling decision from those attributes would act on attacker-chosen values. The content itself remained MAC-bound. asn1.cms.AuthenticatedData now rejects the mismatched pairing when parsing and CMSAuthenticatedDataParser cross-checks the two fields once authAttrs is read. This is a variant of CVE-2026-59642, which bound the content to the MAC for messages that legitimately carry authAttrs, and which does not address this case. This issue also affects Bouncy Castle for Java LTS before 2.73.13, and Bouncy Castle for Java FIPS (BC-FJA) before bcpkix-fips 1.0.13 (1.0.X series), 2.0.13 (2.0.X series) and 2.1.13 (2.1.X series), and bcutil-fips 2.0.8 (2.0.X series) and 2.1.8 (2.1.X series).

In Bouncy Castle for Java before 1.86, neither copy of PKIXCertPathReviewer - org.bouncycastle.pkix.jcajce.PKIXCertPathReviewer nor the legacy org.bouncycastle.x509.PKIXCertPathReviewer - applied X.509 name constraints to the end-entity certificate. checkNameConstraints walked the path with a loop bound of index greater than zero, which is the bound the CA-only steps require, but index zero is the target certificate under the standard CertPath ordering, so the permitted and excluded subtree checks of RFC 5280 sec. 6.1.3 (b) and (c) never ran against the leaf's subject DN or its subjectAltName. A chain whose leaf violated a NameConstraints extension imposed by its own issuing CA therefore reported isValidCertPath() true with an empty error list, while CertPathValidator.getInstance("PKIX", "BC"), which shares no code with the reviewer, rejected the identical chain against the identical trust anchor. An application using the reviewer to make the trust decision rather than for diagnostics alongside a real validation accepted a certificate the constrained CA was never authorised to issue. Both copies now check every certificate in the path including the target, waive the sec. 4.2.1.10 self-issued exemption for the final certificate as sec. 6.1.3 requires, and skip the sec. 6.1.4 (g) constraint-accumulation step for the target. This issue also affects Bouncy Castle for Java LTS before 2.73.13, which carries only the org.bouncycastle.pkix.jcajce copy of the reviewer. It also affects Bouncy Castle for Java FIPS (BC-FJA) before bcpkix-fips 1.0.13 (1.0.X series), 2.0.13 (2.0.X series) and 2.1.13 (2.1.X series).

In Bouncy Castle for Java before 1.86, the opt-in key-size validation on CMS key-transport recipients, org.bouncycastle.cms.jcajce.JceKeyTransRecipient.setKeySizeValidation(true), never ran for a message using RFC 9709 content-encryption key derivation (id-alg-cek-hkdf-sha256). The branch that should have selected the actual content-encryption algorithm carried in the key derivation AlgorithmIdentifier's parameters compared the encrypted-key byte array against the id-alg-cek-hkdf-sha256 object identifier, a comparison between a byte array and an ASN1ObjectIdentifier that is false for every possible input, so the check fell through to a key-size lookup on the outer wrapper OID. That OID identifies a key-derivation construction rather than a cipher and has no registered key size, so the size comparison was skipped entirely. A key-transport EnvelopedData or AuthEnvelopedData whose transported, HKDF-derived content-encryption key did not match the key size of the advertised content-encryption algorithm was therefore accepted even with validation explicitly enabled, silently defeating the only mechanism the API offers for enforcing recovered key size. The recipient now dispatches on the content-encryption AlgorithmIdentifier's algorithm OID, so validation checks the recovered key against the inner content-encryption algorithm. Messages with a matching key size, non-HKDF messages, and recipients that do not enable validation are unaffected. This issue also affects Bouncy Castle for Java FIPS (BC-FJA) before bcpkix-fips 2.0.13 (2.0.X series) and 2.1.13 (2.1.X series).

In Bouncy Castle for Java before 1.86, a truncated OpenPGP encrypted message was accepted with no error reported, and on the SEIPD version 1 path with no integrity check performed at all. RFC 9580 sec. 13.7 permits an implementation to release the cleartext of the fully authenticated chunks when streaming but requires it to indicate a clear error as soon as the truncation is detected, and to report suspect integrity when it discovers malleable ciphertext. The truncation was detected and then discarded: when a message is truncated but the length field of the enclosing packet is left unchanged, BCPGInputStream.PartialInputStream raises an EOFException for the missing ciphertext, and BCPGInputStream.nextPacketTag() reports an EOFException as a clean end of message, so the packet stream above it stopped as though no packets remained. On the AEAD path (SEIPD version 2 and the version 5 AEAD packet), when the literal data packet ended on an AEAD chunk boundary and the consumer read in increments smaller than one chunk, the look-ahead for the packet after the literal triggered the truncated chunk read, so BcAEADUtil and JceAEADUtil never reached the trailing message tag of sec. 5.13.2 that authenticates the total plaintext length; the caller received the plaintext of the fully authenticated chunks, every packet following the literal was silently dropped, and no exception was raised, so a signed and encrypted message read back as a well-formed unsigned one. Every byte released on that path remained individually authenticated, making this a missing truncation error rather than a forgery, and it is a residual of CVE-2026-12817, which closed the same outcome for an attacker who corrects the outer packet length. On the SEIPD version 1 path the consequence was more serious: IntegrityProtectedInputStream verifies the modification detection code from close(), and reached close() only by closing itself when a read of it returned -1, which a truncated message never produces, so PGPEncryptedData.verify() never ran and the recipient was handed CFB-decrypted plaintext on which no integrity check of any kind had been performed. Measured on a message truncated into that shape, 136 distinct single-byte modifications of the ciphertext produced accepted, altered plaintext with no exception raised. Reachability is a property of the message rather than of attacker-supplied input: the AEAD shape held for 3 of 131 consecutive payload lengths measured, and the SEIPD version 1 shape for one payload length in sixteen, at a truncation offset that did not move with the payload length. The low-level API is unaffected, a caller that invokes PGPEncryptedData.verify() directly getting the check regardless, as are consumers reading in increments of a whole AEAD chunk or more. The AEAD decryption streams now re-throw such an EOFException as a plain IOException, which nextPacketTag() does not launder; OpenPGPMessageInputStream.close() now closes its layer's integrity-protected stream itself rather than relying on that stream having seen the end of its data; and IntegrityProtectedInputStream.close() was made idempotent, as java.io.Closeable requires, which that depends on, since the stream is genuinely closed twice on the ordinary path and PGPEncryptedData.verify() consumes the digest state behind it and cannot be run a second time. This issue also affects Bouncy Castle for Java LTS before 2.73.13, on the AEAD route only, as that edition does not ship the high-level OpenPGP API the SEIPDv1 route runs through. It also affects Bouncy Castle for Java FIPS (BC-FJA) before bcpg-fips 1.0.14 (1.0.X series), 2.0.14.1 (2.0.X series) and 2.1.14 (2.1.X series), on the AEAD route only, as those editions do not ship the high-level OpenPGP API.

In Bouncy Castle for Java before 1.86, the raw JCA provider's legacy PBES1 (PKCS#5 scheme 1) and PKCS#12 PBE families ran their password-based key derivation with an iteration count taken from untrusted input without bounding it, so a small input could dictate an arbitrary amount of work before anything could be verified. The AlgorithmParameters implementations (PKCS12PBE and its object identifier aliases, and PBKDF1) accepted any count from an encoded PKCS12PBEParams or PBEParameter, narrowing a value beyond the int range with intValue(), and every Cipher, Mac and SecretKeyFactory in these families derived with whatever count it was given, including one decoded by another provider's AlgorithmParameters, as when javax.crypto.EncryptedPrivateKeyInfo.getKeySpec() decrypts a PKCS#12 PBE-protected private key with BC. Both the parameter parse and the derivations now reject a negative or over-limit count under the org.bouncycastle.pbe.maxiterationcount property (default 10,000,000) that already bounded PBKDF2 (CVE-2026-17508), and the parse rejects a count beyond the int range rather than narrowing it. This issue also affects Bouncy Castle for Java LTS before 2.73.13.

In Bouncy Castle for Java before 1.86, NTRU reduced secret values with the % operator in three helpers whose reference implementations are deliberately division-free, so each reduction was carried out by an integer division whose latency depends on the secret operand. Polynomial.modQ divided by a variable divisor, which a compiler cannot strength-reduce to a multiply the way it can a constant one, so it emitted a division on every call including on the decapsulation path where the dividend derives from the private key; Polynomial.mod3 and NTRUSampling.mod3 divided the secret key polynomials f and g during key generation, the message polynomials r and m during encapsulation, and coefficients recovered during decapsulation. An attacker able to measure that timing can recover information about the NTRU private key. modQ now masks, which is exact because q is always a power of two, and mod3 uses the reference implementation's division-free fold and select; the results are unchanged.

In Bouncy Castle for Java before 1.86, several password-based key derivation entry points ran the KDF with cost parameters taken from the untrusted input being processed, without bounding them, so a small input could dictate an arbitrary amount of work before any password or integrity check could reject it. The affected paths are the RFC 9579 PBMAC1 MAC calculator builders, which took the PBKDF2 iteration count and derived-key length straight out of PBMAC1Params (JcePBMac1CalculatorBuilder, and PKCS12PBEUtils.createPBMac1Calculator reached from PKCS12PfxPdu.isMacValid); the scrypt parallelization parameter p in the PKCS#8 and PKCS#12 cost guards, which bounded only the cost parameter N and the block size r even though the scratch buffer scales with r times p, so the configured memory ceiling could be evaded entirely; the raw JCA PBKDF2 provider (org.bouncycastle.jcajce.provider.symmetric.PBEPBKDF2); and the bcrypt round count read from an encrypted OpenSSH v1 private key's own kdfoptions. Each now bounds the parameter before deriving, in line with the caps already applied elsewhere in the tree, with the OpenSSH round count configurable through the new org.bouncycastle.openssh.maxrounds property. This completes the bounding begun in 1.85 for the PKCS#8 / PBES2 decryptors (CVE-2026-15055). This issue also affects Bouncy Castle for Java LTS before 2.73.13, and Bouncy Castle for Java FIPS (BC-FJA) before bcpkix-fips 1.0.13 (1.0.X series), 2.0.13 (2.0.X series) and 2.1.13 (2.1.X series).

In Bouncy Castle for Java before 1.86, the MLS implementation (org.bouncycastle.mls) holds RFC 9420's uint32 leafindex in a signed int, so a wire value with the top bit set decodes to a negative number. That is a legitimate encoding rather than malformed input, and it must still decode, since the MLS interop test vectors round-trip the full range. GroupKeySet.SecretTree.hasLeaf and Group.validateRemove compared the decoded value directly against the tree's leaf count, and a signed comparison treats any negative int as less than a positive bound, so an out-of-range sender passed the membership check. In the hasLeaf case the SenderData of an unprotected PrivateMessage could then drive LeafIndex.directPath through NodeIndex.parent() arithmetic that never reaches the tree root, growing the resulting node list without bound until the JVM exhausted its heap. A single small message from any current group member could therefore deny service to every other member of the group. Both comparisons now interpret the value as unsigned via Integer.toUnsignedLong, rejecting an out-of-range sender however it was encoded; well-formed leaf indices are unaffected.

Inefficient algorithmic complexity in X.509 distinguished name string conversion (X509Name.ToString and IetfUtilities.ValueToString) in Legion of the Bouncy Castle Inc. bc-csharp before 2.7.0 allows a remote unauthenticated attacker to cause a denial of service through CPU exhaustion via a certificate, CRL, certification request or other structure whose name contains a long attribute value made up of characters that must be escaped, such as commas, or of leading or trailing spaces, because each escaping backslash was inserted into the buffer being scanned, so the work grew quadratically with the length of the value. Applications are exposed when they convert such a name to a string, for example to log or display it, or compare it with IetfUtilities.RdnAreEqual, as PKIX path validation does for directoryName name constraints.

Release of unverified plaintext in the CCM (CcmBlockCipher) and DSTU 7624 CCM (KCcmBlockCipher) AEAD modes in Legion of the Bouncy Castle Inc. bc-csharp before 2.7.0 allows a remote attacker to obtain decryptions of ciphertexts of their choosing via forged messages sent to an application that lets the output buffer of a failed decryption be observed, for example through buffer reuse or logging, because decryption wrote the recovered plaintext into the caller-supplied output buffer before checking the authentication tag and left it there when the check failed. Only decryption into a caller-supplied buffer is affected; methods that return a newly allocated array are not.

Observable discrepancy in the CMS RSA PKCS#1 v1.5 key-transport unwrap (KeyTransRecipientInformation.UnwrapKey) in Legion of the Bouncy Castle Inc. bc-csharp before 2.7.0 allows a remote attacker who holds a captured CMS EnvelopedData message, and who can submit many modified messages to an application that decrypts them with the recipient's RSA private key and reveals how decryption failed, to recover the captured message's content-encryption key and so its content, via a Bleichenbacher-style adaptive chosen-ciphertext attack, because a key-transport ciphertext with invalid PKCS#1 v1.5 padding is rejected during unwrap with a distinct "bad padding in message." CmsException instead of being replaced by a random key, so it can be told apart from a correctly padded ciphertext, which fails only later at content decryption.

Improper verification of cryptographic signature in the attribute certificate path validator (PkixAttrCertPathValidator, also used by PkixAttrCertPathBuilder) in Legion of the Bouncy Castle Inc. bc-csharp before 2.7.0 allows a remote attacker to have a forged X.509 attribute certificate accepted as valid, and so obtain whatever roles or privileges an application grants on the strength of its attributes, via an attribute certificate that names a trusted attribute authority as its issuer but was not signed by it, because the RFC 3281 validation steps check the holder and issuer certification paths, validity period, extensions and revocation status but never verify the attribute certificate's signature with the issuer's public key. Only applications that use these classes to validate attribute certificates are affected.

Loop with unreachable exit condition in Pkcs12Store.GetCertificateChain in Legion of the Bouncy Castle Inc. bc-csharp before 2.7.0 allows an attacker who can supply a crafted PKCS#12 file to an application that loads it and requests a key entry's certificate chain to cause a denial of service, in which the call never returns and consumes CPU and memory until an OutOfMemoryException, via certificates whose issuer links form a cycle, for example two certificates whose AuthorityKeyIdentifier extensions each identify the other's public key. This happens because the chain-building loop stops only when no issuer is found or a certificate links to itself, and keeps no record of certificates already visited. The key-identifier links are followed without checking signatures.

Improper input validation in DHAgreement.CalculateAgreement (MTI/A0 two-pass Diffie-Hellman) in Legion of the Bouncy Castle Inc. bc-csharp before 2.7.0 allows an on-path attacker to make the local party compute an agreed value the attacker already knows, defeating the key authentication MTI/A0 is meant to provide. It also allows a malicious peer to learn the local static private key modulo the small factors of p-1, and to recover it entirely in groups with many such factors. The attack uses a crafted out-of-range or small-order ephemeral value, and works because that value is raised to the static private key without the range and subgroup-membership checks applied to DH public keys. Only applications that call DHAgreement directly are affected.

Observable discrepancy in IesEngine.DecryptBlock in Legion of the Bouncy Castle Inc. bc-csharp before 2.7.0 allows a remote attacker who has captured an IES or ECIES ciphertext, and who can submit modified ciphertexts for decryption under the same key pair, to recover its plaintext via a CBC padding-oracle attack, because in block-cipher mode the engine decrypts the ciphertext and removes its padding before verifying the MAC. A padding failure is therefore reported with a different error message, and without the MAC computation, compared with a MAC failure. Only applications that construct IesEngine directly with a padded block cipher, such as AES in CBC mode with PKCS#7 padding, are affected; stream-mode IES is not.

Memory allocation with excessive size value in the DTLS handshake reassembly (DtlsReliableHandshake, DtlsReassembler) in Legion of the Bouncy Castle Inc. bc-csharp before 2.7.0 allows a remote unauthenticated DTLS peer to cause a denial of service through memory exhaustion via crafted handshake message fragments, because the reassembly buffer for each incoming handshake message was allocated at the 24-bit length declared in the fragment header, without the check against the peer's maximum handshake message size that TLS already applied. A fragment carrying no payload can force an allocation of almost 16 MB, for each of up to 16 pending messages per handshake, before the handshake is authenticated. DTLS servers and DTLS clients are both affected; TLS is not.

Missing cryptographic step in the DSTU 7624 CCM mode implementation (KCcmBlockCipher) in Legion of the Bouncy Castle Inc. bc-csharp before 2.7.0 allows an attacker who can observe encrypted messages of known or chosen content to forge ciphertexts with valid authentication tags, via messages encrypted without associated data. The cause is that the G1 block, which binds the nonce, the message length and the parameter flags into the CBC-MAC, was processed only when associated data was present. Without associated data the tag was a CBC-MAC of the plaintext alone, independent of the nonce. Only applications that use KCcmBlockCipher directly and supply no associated data are affected.

Improper validation of integrity check value in the AES-CCM implementation (CcmParameters and CcmBlockCipher) in Legion of the Bouncy Castle Inc. bc-csharp before 2.7.0 allows an on-path attacker to modify CCM-encrypted content without detection via an AlgorithmIdentifier whose CCMParameters declare an authentication tag (aes-ICVlen) of zero or another length outside the RFC 5084 set, because CcmParameters accepted any value and CcmBlockCipher validated the tag length only when encrypting, so decryption compared a zero-length or very short tag. Affected paths include ParameterUtilities.GetCipherParameters, used by CmsEnvelopedData and CmsEnvelopedDataParser for EnvelopedData encrypted with AES-CCM, and any caller passing an unchecked tag length to CcmBlockCipher for decryption.

In Bouncy Castle for Java FIPS (BC-FJA) before bc-fips 1.0.2.7 (1.0.X series), 2.0.2 (2.0.X series) and 2.1.3 (2.1.X series), sensitive key material held by the AES and DESede engines, the SP 800-90A DRBGs, SymmetricSecretKey and the PBKD and scrypt parameter classes was zeroised on garbage collection by overriding Object.finalize. Finalization runs at an unspecified time and in an unspecified order and is serviced by a single finalizer thread, so where objects carrying a finalizer are allocated faster than that thread retires them the pending-finalization queue grows without bound: disposal falls arbitrarily far behind, which can contribute to an OutOfMemoryError under load, and the key material those objects hold stays resident in the heap for as long as they are queued, defeating the purpose of the zeroisation. The behaviour was not a problem on Java 8 or Java 11; it is later JVMs, on which finalization has been deprecated and progressively de-emphasised, where it becomes one. Disposal of these classes now runs from a java.lang.ref.Cleaner registered in the multi-release jdk1.9 overlay, so on Java 9 and later it no longer depends on the finalizer being scheduled. Bouncy Castle for Java (bcprov) and Bouncy Castle for Java LTS are not affected, as neither implements the finalizer-based zeroisation scheme.

In Bouncy Castle for Java FIPS (BC-FJA) before bc-fips 2.1.3, the native entropy source used on Intel platforms retried the CPU entropy instructions without any bound. RDSEED and RDRAND report failure through their carry flag, and the JNI seeding routine spun re-issuing the instruction for as long as that flag stayed clear, so a persistent failure of the on-chip entropy source - whether from a hardware fault, from the underlying DRBG being exhausted by contention across many cores, or from a hypervisor that does not provide the instruction - left the calling thread looping indefinitely inside the JNI call, where it could be neither interrupted nor timed out. Any operation drawing from the native entropy source could therefore hang, denying service to the application. The retry loops are now bounded (200 attempts for RDSEED and 20 for RDRAND, twice the baselines given in Intel's Digital Random Number Generator software implementation guide), pausing between attempts and, on exhaustion, clearing any partially written buffer and throwing rather than continuing to spin. The clear is performed by an un-elidable memzero, which uses a volatile pointer and an assembly memory barrier so that a compiler cannot optimise the erase away as a dead store. Bouncy Castle for Java (bcprov) is not affected, as it has no native entropy source; the 1.0.X and 2.0.X FIPS series are not affected.

I had the same thought so I asked the project owner and he shared the story :-)

https://github.com/bcgit/bc-java/discussions/2388

Am Mi., 5. Aug. 2026 um 05:24 Uhr schrieb Alan Coopersmith < alan.coopersmith () oracle com>: On 8/3/2026 7:37 PM, Peter Gutmann wrote: Alan Coopersmith <alan.coopersmith () oracle com> writes: It also says the release contains fixes for the following CVEs: Given the quantity and sweeping scope of those, was this the result of some new tool used for code analysis? I'm assuming AI, it sounds like there'd be an interesting backstory to how all of this was turned up. I didn't see anything in the announcements from the Bouncy Castle folks about that.

They do have more info about the CVE's in their wiki, such as: https://github.com/bcgit/bc-java/wiki/CVE%E2%80%902026%E2%80%908763 but I don't see any reference there to how they were found/reported.

On 8/3/2026 7:37 PM, Peter Gutmann wrote: Alan Coopersmith <alan.coopersmith () oracle com> writes: It also says the release contains fixes for the following CVEs: Given the quantity and sweeping scope of those, was this the result of some new tool used for code analysis? I'm assuming AI, it sounds like there'd be an interesting backstory to how all of this was turned up. I didn't see anything in the announcements from the Bouncy Castle folks about that.

They do have more info about the CVE's in their wiki, such as: https://github.com/bcgit/bc-java/wiki/CVE%E2%80%902026%E2%80%908763 but I don't see any reference there to how they were found/reported.

-- -Alan Coopersmith- alan.coopersmith () oracle com Oracle Solaris Engineering - https://blogs.oracle.com/solaris