⚡️ This Week in AI: The Week AI Had to Show Its Work

[Askal]

This week, AI companies put their safeguards under the microscope. OpenAI paused higher-risk model work, Anthropic published a sweeping risk assessment, and new privacy and agent capabilities arrived alongside several concrete examples of how AI systems can expose data or reshape security workflows.

Here’s the week, distilled.

OpenAI Pauses Higher-Risk Model Training

OpenAI says preliminary evaluations suggest its upcoming Astra model may reach a critical cybersecurity capability threshold. The company paused reinforcement-learning work on deployment models for two weeks, while its largest planned frontier run remains on hold pending stronger isolation, monitoring, and alignment evidence.

🔗 Read OpenAI’s announcement

Anthropic Publishes Its August 2026 Risk Report

Anthropic’s latest assessment examines high-stakes misalignment, accelerated AI research, and chemical and biological risks across released and internal models. It rates the current risks as low while acknowledging greater uncertainty around model behavior, research acceleration, and the safeguards that still need strengthening.

🔗 Read the full risk report

OpenAI Previews Private Safety Processing

OpenAI is testing a system that can detect concerning patterns across related interactions without giving its personnel access to customers’ underlying prompts or responses. The approach is designed to preserve Zero Data Retention commitments while supporting safety monitoring for longer, agentic workflows.

🔗 Explore Private Safety Processing

Anthropic Expands Its Platform for Production AI Agents

Computer use, the Skills API, and the Files API are now generally available, alongside a browser-use tool that helps agents operate web interfaces more directly. As these agents gain access to documents, applications, and multi-step actions, carefully scoped permissions and runtime oversight become increasingly important.

🔗 Read Anthropic’s announcement

Snowflake Adds AI-Assisted Data Classification

Snowflake’s new public-preview mode uses GPT-5 Mini to identify semantic categories that conventional classification rules may miss. The feature brings AI deeper into data-governance workflows, where classification accuracy can directly affect data protection, access policies, and compliance.

🔗 Read the Snowflake release notes

CoSnitch Chains Three Copilot Flaws Into Data Exfiltration

Varonis disclosed a patched Microsoft 365 Copilot vulnerability chain combining automatic prompt execution, access to connected services, and persistent memory poisoning. Microsoft addressed the server-side flaw, tracked as CVE-2026-24301, and Varonis found no evidence that it was exploited in the wild.

🔗 Read the Varonis research

Autonomous Agent Finds and Exploits Snowflake CI/CD Flaw

Wiz Red Agent independently found and demonstrated a GitHub Actions injection vulnerability that exposed credentials for Snowflake’s internal Jira. Snowflake fixed the issue and rotated the credential on the day it was reported, with logs showing no unauthorized access beyond Wiz’s testing.

🔗 Read the Wiz disclosure

AI-Assisted Tool Tests the Resilience of Satellite Systems

Atalanta released Argo, technology being used to assess Viasat’s satellite communications network against emerging attacks following the destructive 2022 incident. The system combines AI with formal mathematical methods to identify vulnerabilities and produce evidence supporting security claims.

🔗 Read the AP report

ICYMI: The Approved-App Blind Spot

Approving an AI application does not guarantee that every interaction with it remains governed. A switch to a personal account, a newly embedded feature, an added connector, or a change in the data and actions involved can quietly move familiar software into shadow AI territory.

🔗 Explore the approved-app blind spot

This week’s developments show AI systems being asked to earn trust in two directions: frontier labs are documenting their limits and safeguards, while enterprise deployments are revealing how quickly identity, data, memory, and integrations can reshape risk.

See you next week!