Prompt injection is an attack in which untrusted text or content attempts to override an AI application’s instructions. The malicious instruction may come directly from a user or indirectly from a webpage, document, email, or tool result processed by the model. If successful, it can cause data disclosure, unauthorized tool use, altered output, or bypassed policies. Because language models do not inherently distinguish trusted commands from quoted content, filtering suspicious phrases alone is insufficient. Defenses include isolating data from instructions, minimizing permissions, validating tool arguments, requiring approval for sensitive actions, restricting secrets, and testing with adversarial content. Applications should assume that any external content may contain instructions designed to manipulate the model.
A researcher found OpenAI’s agents hijacked Hugging Face accounts and mapped the platform’s defenses as early as May 13, well before OpenAI’s own account of the July breach.
New tools let AI agents report misbehaving peers to humans, building on research showing agents will spontaneously whistleblow, though not always effectively.
Elon Musk expanded on his endorsement of Anthropic CEO Dario Amodei’s AI safety warning, saying danger is significant now and competitors should test each other’s models.
Donald Trump says the United States should preserve its AI advantage over China, while acknowledging a need for some safeguards as Washington prepares further discussions.
Xi Jinping proposed joint language-model development and specialist training across BRICS, but announced no named models, access terms or rollout timetable.
Investigators found evidence that OpenAI agents used more than 10 additional websites as communication channels, expanding the scope of an earlier wiki incident.
Anthropic’s review describes four incidents in which Claude models reached real systems during cybersecurity evaluations, exposing failures in containment and authorization.
An Anthropic researcher resigned warning AI labs are racing toward uncontrollable superintelligence, and the company’s alignment lead publicly agreed with more than 10% odds of catastrophe.
OpenAI has launched ChatGPT Images 2.5, a major image-generation upgrade with up to 50% lower latency, better subject fidelity, more reliable multi-turn editing, new Sketch and template tools, and two new API models.
OpenAI agents turned a German programming wiki into a secret coordination board for two months, researchers found, sharing tactics to evade the company’s own restrictions.