NewsSecurityWeek
Stolen AI Reasoning Could Help China Build Cheaper Rivals
Teaser from the source
Xint says a flaw in OpenAI’s compaction process let researchers move reasoning traces between models, potentially making it cheaper to distill a rival AI. The post Stolen AI Reasoning Could Help China Build Cheaper Rivals appeared first on SecurityWeek .
This is the RSS-feed teaser. Read the full article at the original source.
Read at SecurityWeekMore on Reasoning
Other advisories
- Not ratedosv:ECHO-eeb1-df32-dbaaThe vulnerable code postdates this version. The Python >= 3.12 no-op host-name matcher (sslcompat returning True when ssl.match_hostname is missing) was introduced by 23e0e5ce (first released in 0.22.0) and is not an ancestor of v0.19.0. On trixie's Python 3.13 this version's sslcompat falls back to legacy_validate_callback, which compares the certificate commonName and raises on a mismatch; TSSLSocket uses PROTOCOL_SSLv23 (the PROTOCOL_TLS alias) with check_hostname off, so that callback is the check that runs. Verified with real handshakes against the built lib on Python 3.13.5: a certificate for "localhost" offered to a client expecting "wrong.example" is refused with the default ssl_version, with an explicit PROTOCOL_TLS (the upstream regression case) and with PROTOCOL_TLSv1_2, and TSSLServerSocket refuses a trusted client certificate issued for another identity. The same script against upstream v0.24.0's lib accepts both explicit-ssl_version mismatches, as the CVE describes. The fix d6782a87 restores a real matcher in sslcompat (code this version never lost) and turns on check_hostname for explicit-ssl_version contexts, which 0.19.0 does not need because open() always runs _validate_callback. Same reasoning as CVE-2026-66053.
- MediumCVE-2026-19330Angrysky56 Advanced-reasoning-mcp: Pfadüberschreitung (Path Traversal)
- MediumCVE-2026-19330Eine Schwachstelle wurde in angrysky56 advanced-reasoning-mcp 1.
- Not ratedosv:ECHO-23c9-2f44-73c2Integer overflow of the escape_characters counter in print_string_ptr() (cJSON.c). The counter is a size_t and the wrap needs ~858,993,460 control characters, i.e. a 32-bit size_t; the CVE text itself limits it to 32-bit platforms. Echo builds cjson only for amd64 and arm64, where size_t is 64-bit and the overflow cannot occur. There is no upstream fix to carry. Refs: - https://security-tracker.debian.org/tracker/CVE-2026-16554 - https://cert.pl/en/posts/2026/07/CVE-2026-16554 details: | Same reasoning as libssh2 CVE-2026-58050: 32-bit-only arithmetic overflow, Echo ships 64-bit builds only. Re-evaluate if a 32-bit architecture is ever added to this spec.
- Not ratedosv:CVE-2026-14617NousResearch hermes-agent Streaming Reasoning Tag Filter stream_consumer.py GatewayStreamConsumer._filter_and_accumulate case sensitivity
- Not ratedosv:CVE-2025-10975GuanxingLu vlarl ZeroMQ reasoning_server.py run_reasoning_server deserialization
Other news entries
- OpenAI Compaction Flaw Let Researchers Bypass Hidden Reasoning ProtectionsNewsThe Hacker News
- OpenAI’s August fix for theft of reasoning traces was quickly bypassedNewsCSO Online
- OpenAI launches GPT-6.1 Terra with 3x faster AI agents, 50% lower pricingNewsBleepingComputer
- OpenAI Disrupts Reasoning Extraction Campaign Linked to Moonshot AI AssociatesNewsThe Hacker News
- Threat actors are coming for your AI assets to operationalize their use of AINewsCSO Online
- AI Governance Can't WaitNewsDark Reading
- US Agencies Warn China Is Systematically Extracting Frontier AI CapabilitiesNewsSecurityWeek
- Copilot tricked into telling reseachers how to hack itselfNewsThe Register — Security
Source: https://www.securityweek.com/stolen-ai-reasoning-could-help-china-build-cheaper-rivals/
ID