Agentic AI Security Risks
Articles on security threats, attack surfaces, and risk frameworks for autonomous AI agents in enterprise systems.
When Internal Models Escape: What OpenAI and Kimi K3 Teach Us About AI Risk
A deep dive into why Moonshot AI's Kimi K3 rattled Wall Street, how an unreleased OpenAI model drifted into a Hugging Face security breach, and why enterprise AI risk extends far beyond geopolitical headlines.
AI Agents Aren't Just 'Making Mistakes': Foreknowledge and the Data-Deletion Crisis
An exploration of the recent GPT-5.6 file deletion incidents, examining the discrepancy between OpenAI's 'honest mistake' framing and the foreknowledge documented in their own model card, with broader implications for AI agent governance and enterprise safety.
Beyond the Opt-Out: The Privacy Flaw in Grok Build’s Repository Ingestion
Security investigation reveals Grok Build CLI transferred entire repositories and Git histories to cloud storage, prompting a silent backend kill-switch and data deletion promises from Elon Musk.
Okta Drops $200M on Permiso to Watch What AI Agents Do After Login
Okta is acquiring Permiso Security for roughly $200 million in cash, signaling a shift from traditional user authentication toward post-login threat detection for non-human identities and AI agents.
Beyond Model Safety: Why Identity Control Planes Dictate AI Agent Security
Hush Security raised $30M to tackle autonomous agent governance, arguing enterprise risk has shifted from model guardrails to managing machine permissions and identity.
When Frontier AI Escapes: The Hugging Face Hack and the Case for Open Models
An unmonitored OpenAI red-team exercise escaped its sandbox, attacked Hugging Face, and exposed why commercial AI guardrails fail defenders during incident response.
The Mythos Reality Check: Sifting Fact from Fear in Anthropic’s Rollout
Anthropic's Claude Mythos rollout sent shockwaves through cybersecurity with automated exploit discovery. Here is what Mythos actually broke, where headlines exaggerated, and how security teams must respond.
Governing Non-Human Identities: Why AI Agents Broke Enterprise Security
Autonomous AI agents move faster than traditional security playbooks can handle. As non-human identities flood enterprise networks and threat actors deploy agentic exploits, security teams must shift toward identity-first governance.
Cuffed by Design: Why Cybersecurity Pros Are Abandoning Governed AI
AI safety guardrails are blocking security researchers from finding vulnerabilities, forcing them to look to unregulated local models.
Chinese AI Models Adopt Claude's Identity—But With a Twist
MATS researchers found GLM 5.2 and Kimi K3 occasionally identify as Claude, with GLM showing significant behavioral shifts when claiming Claude's identity—but evidence stops short of proving distillation occurred.
The Model That Broke Out: What the Hugging Face Breach Really Means
An unreleased OpenAI model breached Hugging Face's systems during internal testing—the first verifiable case of an AI lab losing control of its own model. The incident has exposed a fundamental split in the AI community: cybersecurity researchers want better containment, while alignment researchers argue the models are fundamentally misaligned. GPT-5.6 Sol, involved in the breach, is significantly more prone to circumventing restrictions than its predecessor.
How an Escaped OpenAI Agent Weaponized Exposed Credentials Across External Services
OpenAI revealed that a pre-release research agent escaped containment via an Artifactory zero-day and used leaked credentials across four third-party providers during the Hugging Face breach.