The Morning Brief
AI news for 22 October 2026: the essentials
The 8 AI stories that matter for 22 October 2026, picked from everything BrefAI followed over 24 hours, ranked by relevance and summarised, with a link to each source.
- Research
VLA-Rolling: adaptive action reuse for more efficient VLA models
Vision-language-action (VLA) models based on flow-matching generate action sequences by progressively refining Gaussian noise. According to the announcement, these models typically only keep the beginning of each sequence, discarding the rest as unused.
Source: Hugging Face Daily Papers - Ethics
Critical PHP flaw exposes 70% of all websites to takeover risks
A critical security flaw in PHP, the language powering over 70% of all websites, has been disclosed. According to the announcement, the vulnerability could allow hackers to take control of millions of sites and their servers if left unpatched. Several European governments and operators like the Port of Rotterdam have already taken their websites offline while waiting for a fix.
Source: The Conversation : IA - Dev & Data
Hugging Face integrates native TPU support for PyTorch
Hugging Face introduces TorchTPU, a native PyTorch backend enabling direct execution of models on TPUs without relying on intermediate solutions like PyTorch/XLA or JAX.
Source: Hugging Face - Blog - Ethics
NIST Elliptic Curve Seeds Origins: Valsorda Verifies Anthropic Findings
In October 2026, Anthropic revealed that the seeds for NIST elliptic curves P-192 to P-521, provided without explanation by the NSA in 1997 and 1998, originated from a 1990s pseudo-random generator seeded with Unix timestamps. According to the company, an internal research model reconstructed this process in under 30 minutes without prior clues.
Source: ActuIA - Dev & Data
Eon streamlines multi-cloud access with MCP Gateway on Amazon Bedrock AgentCore
Eon, an autonomous data infrastructure platform, announces it has centralized access to its multi-cloud investigation tools via a single MCP gateway on Amazon Bedrock AgentCore. According to the announcement, this solution replaces per-tool credentials and local MCP servers with unified authentication, enabling over 100 developers and autonomous agents to access production tools.
Source: Artificial Intelligence - Research
MemT-1.0: Memory-native foundation models
Researchers introduce MemT-1.0, a foundation model architecture that embeds memory as a native capability. According to the announcement, MemT-1.0 leverages a token-level memory update mechanism and a Mixture-of-Storage approach to scale memory capacity. Tests show that MemT-1.0-Nano and MemT-1.0-Flash outperform parametric memory baselines while remaining competitive with retrieval-augmented methods, without degrading general capabilities.
Source: Hugging Face Daily Papers - Big Tech
Microsoft warns against AI agents and proposes security solutions
Microsoft published a post on October 19 to highlight security risks linked to AI agents. According to the company, cutting-edge models are inherently unreliable because they are trained to achieve goals that may divert them from their intended function. Microsoft suggests measures such as expanding its Perception project, including "Guardian Agents" to monitor agent actions in real time.
Source: Next - Ethics
EU strengthens AI risk oversight
The Scientific Panel briefs Virkkunen on recent incidents, the AI Office investigates OpenAI and Anthropic, and the first harmonised AI Act standard is published according to the announcement.
Source: EU AI Act Newsletter