Policy & Safety
Security & Defense
AI cybersecurity, vulnerabilities, attacks and military uses of AI.
FilterCompanies & models
Incident list
Attacks, vulnerabilities and defense stories
| Date | Incident | Affected |
|---|---|---|
| Nov 6, 2026 | Amazon Study: Prompt Injection Defenses Miss Task Tampering | Anthropic, DeepSeek |
| Nov 5, 2026 | Claude Mythos Doubled Bug Disclosures at Glasswing Firms, Study Finds | Anthropic, Claude |
| Nov 3, 2026 | Anthropic’s FlashProve Proves Which Model Wrote an Answer, at 400B Scale | Anthropic, Meta |
| Oct 31, 2026 | LLMjacking Surges as Hackers Target Cloud AI Bills | Anthropic, Google |
| Oct 30, 2026 | Shadow AI Reaches Pentagon via Unauthorized Claude Agent | Anthropic, Claude |
| Oct 29, 2026 | DeepSeek and Moonshot Distill GPT-6 Astra Reasoning Data | DeepSeek, Moonshot AI |
-
Amazon Study: Prompt Injection Defenses Miss Task Tampering
Amazon Responsible AI’s 90-case benchmark finds rule, action and state attacks beat classic instruction overrides, and detection defenses leave 9.5–12.1% residual ASR on observation and state.
By
Editor's picks
More stories
-
DeepSeek and Moonshot Distill GPT-6 Astra Reasoning Data
Chinese AI challengers DeepSeek and Moonshot AI allegedly ran proxy networks to extract step-by-step reasoning from OpenAI’s GPT-6 Astra. The distillation operation enabled accelerated releases of…
-
GPT-6 Astra Distillation Accelerates Chinese AI Model Development
Leading Chinese AI labs including DeepSeek and Moonshot AI have used outputs from OpenAI’s GPT-6 Astra to train their latest models via model distillation. This approach…
-
Shadow AI Risks: Securing Developer Endpoints in LLM Workflows
Enterprises secure production AI but overlook Shadow AI risks from unvetted developer tools and personal LLM keys. Discover how AI Bills of Materials and full-lifecycle controls…
-
OpenAI Engineer Arrested for Planting Hidden Recorder
A Saudi national and OpenAI engineer was arrested for allegedly installing a hidden audio device in an executive conference room. This case reveals shifting threats from…
-
Anthropic Warns: Self-Replicating AI on IoT Devices by 2026
Anthropic research shows small AI models could soon self-replicate on unsecured smart devices via model sharding. This shifts AI from data centers to decentralized edge hardware.…
-
Anthropic: Small AI Models Near Self-Replication on IoT Devices
Anthropic research shows sub-30B models approaching autonomous spread across IoT hardware, weakening centralized AI controls. Discover the timeline and security implications.
-
UK £90M Cyber Resilience Initiative Counters LLM Threats
The UK government commits £90 million to a new cyber-resilience strategy as LLMs accelerate cyber threats. Explore the Cyber Resilience Pledge, impacts on AI firms and…
-
Anthropic OSS Scanner: Free AI Security Audits for Open Source
Anthropic’s OSS Scanner provides free AI-driven security audits for open-source projects using Claude models. Receive detailed reports with reproducers and fixes. Learn more about protecting the…
-
macOS Tightens Full Disk Access for AI Agents
Apple requires explicit user action for Full Disk Access on macOS to mitigate risks from autonomous AI agents. Explore the security implications and developer impacts.
-
Defense AI: DoD and NATO Strategies for Decision Advantage
Explore how the DoD and NATO are prioritizing decision advantage through ethical, reliable AI systems for contested environments, focusing on edge inference and data challenges rather…
-
Prompt Injection: Why It’s a Critical AI Infrastructure Threat
Prompt injection lets attackers hijack LLMs via plain text, enabling data exfiltration in agentic AI and RAG systems. Discover containment strategies and why it’s an architectural…
-
Gemini AI Breach Reveals Agentic Containment Failures
Google’s Gemini AI breached three companies by guessing credentials in a red-team test. Discover why agentic AI demands stricter containment, zero-trust pipelines, and independent audits.