Fixes of the review of the session of 1 and 2 October in internal/cms and
internal/der:
- Object identifiers are compared by the bytes of their DER: an arc of 2^31
or more no longer makes an attribute that decides nothing fail the
signature (F1), and an algorithm with one is outside the table (F5, S1),
as spec v0.11 says.
- A SET OF may repeat an element, as X.690 allows: a TSA that sends its
certificate twice no longer gives S2. Two copies of a certificate are one.
- Certificates are read with a profile of their own instead of
encoding/asn1 and crypto/x509, field by field, so that a second
implementation can read them the same: version 3, the fields in order,
names of non-empty SETs, times of validity in DER without a fraction,
extensions without repetition, and the text of a name only from
UTF8String, PrintableString, IA5String, TeletexString in ASCII and
BMPString without surrogates, nothing removed from it. A certificate that
breaks the profile decides nothing unless a SignerInfo names it. The
holder is givenName and surname before commonName, which in the
certificates of the FNMT carries the NIF; the issuer is its commonName or
its organizationName.
- The key: RSA with NULL parameters, exactly a modulus and an exponent, the
modulus odd; EC only uncompressed on P-256, P-384 and P-521.
- A key of another scheme than its algorithm is invalid (F2) and not
outside the table (F5), as step 3 of 29.10 says; a messageImprint of
another length is S3; the crls of a token decide nothing.
- DER: UTCTime and GeneralizedTime in their forms of X.690, a date that
exists, and the millis and micros of accuracy as minimal INTEGERs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The TSTInfo is read field by field in DER, with accuracy from zero and
millis and micros from 1 to 999, genTime in UTC with Z, no default written
and nothing after the last field. The ContentInfo and the SignerInfo must be
SEQUENCEs, a SignerInfo version must match its sid, an attribute needs a
value and is counted by attribute and not by value, a signing-certificate
beside the v2 decides nothing, PSS parameters come in order without the
trailer, and der.Check refuses the end of contents and the universal tags
the profile does not use.
The writer signs before L is fixed: write asks prepare for the final L, so
the area grows to 64 KiB only when what was signed does not fit and LargeArea
allows it, and nobody signs twice for it. Typed nils are nil, the exclusions
are checked before a file is read, Encrypt refuses the signing options, and
EvaluateSecurityIn gives X if a parser panics. The issuer of a certificate is
filtered like its holder.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
internal/der checks that bytes are one element of DER. internal/cms reads
the detached CMS signature of spec v0.11 29.10 and the RFC 3161 token of
29.11, in the order of the spec, with the closed table of algorithms (RSA
PKCS 1 and PSS of 2048 to 4096 bits, ECDSA on P-256, P-384 and P-521,
SHA-2), with the standard library only. A certificate is read with
encoding/asn1, so that a key of a curve Go lacks makes a signature "not
verifiable" and not malformed. internal/cms/cmstest builds them for tests.
Not wired into capsule yet.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
dev