Perplexity wins appeal, Amazon ban on Comet lifted

Perplexity overturns Amazon ban on Comet AI shopping agent

The US Court of Appeals for the Ninth Circuit overturned a court order that had banned Perplexity AI’s Comet AI agent from Amazon.com. The appeals court ruled that the injunction was improper because Comet does not directly access Amazon’s computers; instead, users access the site with the agent’s assistance. Perplexity can now resume sending screenshots of Amazon pages users accessed to its servers to complete requested tasks, as that does not qualify as accessing Amazon’s servers, the opinion said.

Perplexity Overturns Amazon Ban on AI Shopping Bot on Appeal (1) →

Linux Foundation publishes first voluntary AI incident-disclosure framework at Black Hat

SAFE proposes six core design choices: confidential reporting rather than mandatory disclosure, separation of learning from enforcement, inclusion of near-misses and AI agent misbehaviors in scope, full-stack scope covering models and tools, machine-readable defensive outputs, and independent governance. The framework is modeled on NASA’s Aviation Safety Reporting System and FS-ISAC. OSAIA grew from approximately 37 founding organizations on July 27 to more than 120 members, with Amazon joining on publication day and contributing two open-source projects. OpenAI, Google, and Anthropic are absent from OSAIA. Turing Award winner Yoshua Bengio called the Hugging Face breach deeply concerning and a wake-up call, warning the current trajectory will likely lead to more autonomous cyberattacks.

AI Security Gets Its First Voluntary Incident-Disclosure Framework at Black Hat →

JFrog finds 54 of 55 vulnerability reports are AI fabrications

The incident exposes a broken verification pipeline: no step in the CVE process requires anyone to reproduce a bug, and NIST paused deep analysis in early 2024 due to a flood of submissions, leaving a backlog of over 27,000 unprocessed reports. Security teams increasingly point AI coding agents at code to fix reported flaws, meaning a fabricated CVE could lead an agent to hunt for a nonexistent vulnerable function and change working code to patch a phantom problem. JFrog’s Afek Berger framed it as an asymmetry: generating a convincing fake costs almost nothing, while verifying one requires reading source, building the version, and running the proof of concept.

Someone posted 55 vulnerabilities in days. 54 were AI-invented. One still scored a perfect 10. →