OpenAI’s GPT-5.5 is as Good as Mythos at Finding Security Vulnerabilities
2026-05-14T07:23:45Z•0125154d1e5b23f12660ed2d2041790fe33e34f05a6416979cd2b9be837b1de8
AF_ALGCVE-2026-31431DarkSwordIOMMUcopy.failfacial-recognitiongpu-exploitiOS-exploit-chaininsider-tradinglinux-kernelllm-steganographylocal-privilege-escalationnvidiaopenai-gpt-5.5page-cachepolymarketproof-of-conceptransomwarerowhammersmart-glassessplicestate-sponsoredvulnerability-discoveryzero-day
What happened
Collection of security news: The UK AI Security Institute found OpenAI’s GPT-5.5 comparable to Anthropic’s Claude Mythos at identifying vulnerabilities. A severe Linux kernel local privilege escalation—copy.fail (CVE-2026-31431)—was disclosed with a working PoC: it abuses the kernel crypto API (AF_ALG) and splice() to write into a file’s page cache across many distributions while evading checksum-based monitoring. Other notable items include LLM text-in-text steganography research, new GPU rowhammer attacks against NVIDIA Ampere cards that can yield full host compromise when IOMMU is disabled,
Why it matters
A reviewed impact interpretation has not been published for this record.
Evidence and limitations
- Source ID
- schneier_blog
- Record identifier
- 0125154d1e5b23f12660ed2d2041790fe33e34f05a6416979cd2b9be837b1de8
- Enrichment time
- 2026-05-14T07:23:45Z
- AI-assisted enrichment
- Yes
This record may overlap with other records. Its enrichment can be incomplete or wrong, and machine assistance was used. Validate consequential decisions against the linked source and your own environment.