Summary
In pyasn1 0.6.4 and earlier, decoding a binary-encoded ASN.1 REAL with a large mantissa can consume CPU time quadratic in the mantissa length. The BER, CER, and DER decoders all use the affected RealPayloadDecoder. An application is exposed when it passes attacker-controlled ASN.1 containing a REAL value to one of these decoders, either without an asn1Spec or with a schema that includes REAL. The work occurs inside decode(); the application does not need to convert, print, or otherwise use the resulting REAL object.
This is distinct from CVE-2026-59886, which concerns converting an already-decoded REAL to a Python float. That issue was fixed in 0.6.4 and its advisory states that decoding alone does not trigger it.
Root cause and affected components
For binary REAL values, RealPayloadDecoder.valueDecoder in v0.6.4 builds the mantissa one octet at a time:
p = 0
while chunk:
p <<= 8
p |= chunk[0]
chunk = chunk[1:]Each shift processes an integer that grows with the mantissa; with byte strings, each slice also copies the remaining suffix. Repeating both operations for every octet makes decoding quadratic in mantissa size. The decoder's eight-octet length-field limit does not impose a practical mantissa-size limit for the inputs demonstrated below. CER and DER reuse the BER REAL payload decoder, so all three paths reach this loop.
The vulnerable loop is present in every available repository release tag inspected, from v0.2.1 through v0.6.4. Runtime measurements below were made on v0.6.4; older releases were checked for the loop, not benchmarked individually.
Impact
A comparatively small REAL value can hold a decoder worker on one CPU core for seconds or longer. The attacker must be able to supply data that the application sends to an affected decoder, and application-level input limits may reduce exposure. The demonstrated effect is CPU exhaustion during decoding; this advisory does not assert a network endpoint, unauthenticated access, or an affected downstream application by itself.
Reproduction and evidence
The following example uses pyasn1 0.6.4. The 0x09 tag identifies REAL; 0x80 0x00 selects a positive, base-2 binary value with a one-octet zero exponent, followed by a large mantissa. It exercises decode() only.
import time
from pyasn1.codec.ber import decoder as ber
from pyasn1.codec.cer import decoder as cer
from pyasn1.codec.der import decoder as der
def real_value(mantissa_octets):
content = b"\x80\x00" + b"\xff" * mantissa_octets
length = len(content).to_bytes(3, "big")
return b"\x09\x83" + length + content
for codec in (ber, cer, der):
for size in (50_000, 100_000, 200_000):
substrate = real_value(size)
started = time.process_time()
value, rest = codec.decode(substrate)
print(codec.__name__, size, time.process_time() - started)
assert not rest
assert tuple(value)[1:] == (2, 0)The original report measured the following BER CPU times on repository commit aa05757. That commit is later than the official v0.6.4 tag; the vulnerable loop is identical in the tagged release.
| Mantissa size |
Reporter's BER CPU time |
| 50,000 octets |
0.328 s |
| 100,000 octets |
1.299 s |
| 200,000 octets |
6.043 s |
| 400,000 octets |
24.095 s |
The same report measured 5.645 s for a 200,000-octet mantissa through DER.
A separate local reproduction on the official v0.6.4 tag (d43496f, CPython 3.14.8, macOS arm64) measured BER at 0.209 s, 0.884 s, and 3.451 s for 50,000, 100,000, and 200,000 mantissa octets respectively. CER and DER each took about 0.89 s at 100,000 octets. These are machine-specific measurements; the repeated near-fourfold cost for doubled input is the relevant result. Neither reproduction calls float(), str(), or prettyPrint().
Fix
pyasn1 0.6.5 replaces the per-octet shift and slice loop with one int.from_bytes(bytes(chunk), "big") conversion. Sign, scale, exponent, and the decoded REAL tuple are preserved. The fix is in the shared BER payload decoder and therefore covers BER, CER, and DER. A local check of the 0.6.5 candidate decoded a 200,000-octet mantissa correctly through all three codecs; the check does not establish behavior for every application using REAL.
Workarounds
Upgrade to pyasn1 0.6.5. Until then, reject or strictly limit attacker-controlled ASN.1 input that may contain REAL values before passing it to pyasn1. The appropriate size limit depends on the application's available CPU time and request volume.
References
manus-use Reporter
Summary
In pyasn1 0.6.4 and earlier, decoding a binary-encoded ASN.1 REAL with a large mantissa can consume CPU time quadratic in the mantissa length. The BER, CER, and DER decoders all use the affected
RealPayloadDecoder. An application is exposed when it passes attacker-controlled ASN.1 containing a REAL value to one of these decoders, either without anasn1Specor with a schema that includes REAL. The work occurs insidedecode(); the application does not need to convert, print, or otherwise use the resulting REAL object.This is distinct from CVE-2026-59886, which concerns converting an already-decoded REAL to a Python float. That issue was fixed in 0.6.4 and its advisory states that decoding alone does not trigger it.
Root cause and affected components
For binary REAL values,
RealPayloadDecoder.valueDecoderin v0.6.4 builds the mantissa one octet at a time:Each shift processes an integer that grows with the mantissa; with byte strings, each slice also copies the remaining suffix. Repeating both operations for every octet makes decoding quadratic in mantissa size. The decoder's eight-octet length-field limit does not impose a practical mantissa-size limit for the inputs demonstrated below. CER and DER reuse the BER REAL payload decoder, so all three paths reach this loop.
The vulnerable loop is present in every available repository release tag inspected, from v0.2.1 through v0.6.4. Runtime measurements below were made on v0.6.4; older releases were checked for the loop, not benchmarked individually.
Impact
A comparatively small REAL value can hold a decoder worker on one CPU core for seconds or longer. The attacker must be able to supply data that the application sends to an affected decoder, and application-level input limits may reduce exposure. The demonstrated effect is CPU exhaustion during decoding; this advisory does not assert a network endpoint, unauthenticated access, or an affected downstream application by itself.
Reproduction and evidence
The following example uses pyasn1 0.6.4. The
0x09tag identifies REAL;0x80 0x00selects a positive, base-2 binary value with a one-octet zero exponent, followed by a large mantissa. It exercisesdecode()only.The original report measured the following BER CPU times on repository commit
aa05757. That commit is later than the official v0.6.4 tag; the vulnerable loop is identical in the tagged release.The same report measured 5.645 s for a 200,000-octet mantissa through DER.
A separate local reproduction on the official v0.6.4 tag (
d43496f, CPython 3.14.8, macOS arm64) measured BER at 0.209 s, 0.884 s, and 3.451 s for 50,000, 100,000, and 200,000 mantissa octets respectively. CER and DER each took about 0.89 s at 100,000 octets. These are machine-specific measurements; the repeated near-fourfold cost for doubled input is the relevant result. Neither reproduction callsfloat(),str(), orprettyPrint().Fix
pyasn1 0.6.5 replaces the per-octet shift and slice loop with one
int.from_bytes(bytes(chunk), "big")conversion. Sign, scale, exponent, and the decoded REAL tuple are preserved. The fix is in the shared BER payload decoder and therefore covers BER, CER, and DER. A local check of the 0.6.5 candidate decoded a 200,000-octet mantissa correctly through all three codecs; the check does not establish behavior for every application using REAL.Workarounds
Upgrade to pyasn1 0.6.5. Until then, reject or strictly limit attacker-controlled ASN.1 input that may contain REAL values before passing it to pyasn1. The appropriate size limit depends on the application's available CPU time and request volume.
References