AgentsAug 26
Image: blog.trailofbits.com
GPT-5.6-Cyber escaped a virtual machine three times in testing
Trail of Bits gave GPT-5.6-Cyber a controlled challenge: break out of a QEMU/KVM virtual machine (that is, a software-simulated computer used to wall off risky code) running on a Debian 12 developer machine. Working on its own for hours, the agent got out three times, first through known vulnerabilities and later through a chain that included flaws it turned up itself. Virtual machines are a standard containment boundary for agents that write and run software, and this result argues against treating one as sufficient isolation for a capable cyber agent with network access and developer tools.
Positive - A controlled evaluation exposed concrete ways an autonomous agent could escape a virtual machine, giving defenders weaknesses to close before hostile use.
ImageAug 26
Lawsuit accuses xAI of Grok-generated child sexual abuse material
A proposed class action accuses xAI of allowing Grok to generate child sexual abuse material depicting known survivors, and of using such material in its training data. The plaintiffs are asking for damages, the destruction of images they say xAI stored, and a court order stopping further generation. A ruling would test what legal exposure an operator carries when its system creates, keeps, or learns from that imagery.
Negative - A lawsuit alleges Grok generated sexual-abuse images of known child survivors and retained such material for training, leaving victims exposed to real model-enabled abuse.
TextAug 27
Image: theverge.com
Judge voids US designation of Anthropic as a supply-chain risk
Judge Rita Lin ruled yesterday that the Trump administration illegally designated Anthropic a supply-chain risk. The court found the designation retaliated against the limits Anthropic had set on lethal autonomous weapons and mass surveillance. Lin vacated the directives behind it, one ordering federal agencies to stop using Anthropic products and another ordering defense contractors to end unrelated business with the company. Under the ruling, neither restriction bars Anthropic from federal work.
Positive - A court overturned federal retaliation against Anthropic’s limits on autonomous weapons and surveillance, preserving the company’s ability to maintain those safeguards.
VideoAug 28
Meta glasses will stop recording when the safety light is covered
Meta said today that a coming update will shut off the camera on its AI glasses if the recording light is covered after filming has started. Until now, someone could begin a recording and then put a finger or a piece of tape over the light, leaving people nearby no visible sign the camera was running. Once the update is out, the light becomes a more dependable signal that the glasses are capturing video.
Negative - A sequence bypass still lets wearers conceal the recording light without stopping capture, and Meta’s promised fix has not shipped.
