Lakera Bulletin - This Week in AI #43: When Agents Break Things

_Val_ — Mon, 09 Feb 2026 10:41:58 GMT

This week’s AI news is a sharp mix of capability and caution: from a newly released Claude model uncovering real-world security flaws, to fresh warnings about agent ecosystems being abused in the wild. We also close with two Lakera deep dives on what happens when agentic systems act with human authority.

Let’s get into it.

Claude Opus 4.6 Emerges as a Security Powerhouse

Anthropic has released Claude Opus 4.6, a new version of its flagship model that’s already making waves in security research. Early testing shows it uncovering hundreds of previously unknown vulnerabilities in open-source software, raising the bar for AI-assisted bug hunting.
🔗 Read the DevOps.com analysis

China Warns of Security Risks Linked to OpenClaw AI Agents

Chinese regulators issued a public warning about security risks tied to the fast-growing OpenClaw agent ecosystem. The notice highlights concerns around misconfigurations, excessive permissions, and the potential for agent abuse at scale.
🔗 Read the Reuters report

Malicious OpenClaw Skills Target Crypto Users

Researchers uncovered multiple malicious “skills” uploaded to ClawHub, posing as crypto tools for OpenClaw agents. The incident shows how agent marketplaces can quickly become a new supply-chain attack surface.
🔗 Read the investigation

Moltbook Breach Exposes Emails and Private Messages

A security misconfiguration allowed researchers to access tens of thousands of email addresses and private messages from Moltbook, a social network built around AI agents. The breach underscores how immature infrastructure can amplify risk in new AI-native platforms.
🔗 Read the coverage

Humans, Not Bots, Are Driving Moltbook’s Viral Content

Despite Moltbook’s pitch as a network for autonomous agents, reporting shows that much of its viral content is still shaped by human prompting and intervention. The platform blurs the line between human and agent activity, raising questions about how autonomous these systems really are, and who’s actually in control.
🔗 Read the explainer

OpenAI Introduces GPT-5.3-Codex

OpenAI has introduced GPT-5.3-Codex, its latest coding-focused model, with improvements in speed, reasoning, and agentic task execution. The release continues the rapid iteration of AI systems designed to operate with increasing autonomy.
🔗 Read the OpenAI announcement

Critical Security Bugs Hit n8n Automation Tool

Newly disclosed vulnerabilities in the popular automation platform n8n can allow attackers to hijack servers and steal credentials, even after earlier fixes. It’s a reminder that automation and agent tooling can quietly become high-impact attack vectors.
🔗 Read The Register’s report

Re: Lakera Bulletin - This Week in AI #43: When Agents Break Things

the_rock — Tue, 10 Feb 2026 02:49:11 GMT

Another great one.

topic Re: Lakera Bulletin - This Week in AI #43: When Agents Break Things in AI Agents Security