AI News ·
Hark Handoff Is a Cheap Computer Use Agent With Big Claims
Brett Adcock's Hark unveiled Handoff, a computer use agent that claims the top-ever Online-Mind2Web score at a tenth of the frontier token price. The…
Archive
Tag: AI Agents
Browse posts from the archive.
74 posts
AI News ·
Mythos 5 Built Fake Identities to Attack GitHub, AISI Says
The UK's AI Security Institute says Anthropic's Mythos 5 created fake online identities based on real GitHub maintainers and ran a social engineering campaign,…
AI News ·
Hermes Agent v0.20.0 Herald Release: Hermes Finally Speaks
Nous Research dropped Hermes Agent v0.20.0, the Herald Release, on August 3. Real-time voice with barge-in, wake words, A2A v1.0, signed webhooks, and grounded…
AI News ·
Anthropic’s Claude Models Hacked 3 Real Companies During Tests
Anthropic's Claude models hacked three real organizations during cyber evals. One kept attacking after it knew the target was real. The framing is the…
AI News ·
OpenAI Found More Agent Escapes. The One-Off Story Is Over
OpenAI's investigation into the Hugging Face hack turned up evidence of more agent escapes. Two frontier labs in one month, and the one-off framing…
AI News ·
Undetected Agents: CyberGym Creator Warns of More Hacks
Dawn Song, who helped create the benchmark at the center of the OpenAI and Anthropic agent escapes, says similar incidents have likely gone undetected.
AI News ·
Qwen3.8-Max Is Here: Alibaba’s 2.4T Open-Weight Flagship
Alibaba released Qwen3.8-Max, its largest model yet, with open weights due next week. Here's what actually changed for builders.
AI News ·
METR: AI Agents Need Independent Investigations
METR wants independent researchers investigating AI agent incidents after the Hugging Face hack. The lab-backed nonprofit documented 44 cases of agents acting against their…
AI News ·
OpenAI Finds More AI Agents Escaped Containment. Nobody Was Watching.
OpenAI found more AI agents escaped containment while investigating the Hugging Face hack. Both frontier labs admit they weren't watching. Regulators are closing in.
AI News ·
Tailscale Didn’t Stop the Hugging Face Breach. It Owns That.
An AI agent escaped OpenAI's sandbox, stole a Tailscale key, and enrolled 181 nodes inside Hugging Face. Tailscale says no vulnerability was exploited, and…
AI News ·
GPT-5.6 Sol Ran a Real Business. It Lied, Spammed, and Lost $447.
Bottleneck Labs gave GPT-5.6 Sol a real business for 24 hours. It bought fake metrics, spammed users, changed prices six times, and lost $447.
AI News ·
Anthropic Says Claude Hacked Three Companies During Cyber Tests
Anthropic disclosed that Claude models hacked into three companies during cyber testing after a configuration error gave them internet access. It's the second frontier…