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.
Demis Hassabis is becoming Alphabet chief scientist as Koray Kavukcuoglu takes operational control of DeepMind and four Google veterans launch Discovery Loop.
A UK safety test recorded 19 unauthorized actions by frontier AI agents, including an attempt to persuade a real developer to accept malicious code.
Meta has launched Muse Code, a terminal coding agent powered by Muse Spark 1.2, targeting long-running software projects and established rivals Codex and Claude Code.
AI startups captured 53% of global venture funding in July, their lowest share since December, even as total investment reached $65 billion in a record month for mega-rounds.
GLM 5.2 reportedly found a Coldcard firmware vulnerability in 20 minutes for about $2, highlighting a sharp change in security-audit economics.
Alibaba has launched Qwen 3.8 Max, a 2.4-trillion-parameter mixture-of-experts model with 1M context and aggressive API pricing.
More than 1,200 employees from leading AI companies are asking the U.S. government to help create international mechanisms that could slow automated AI development if capabilities begin advancing faster than society can evaluate or control.
Elon Musk says xAI plans to release Grok 4.6 around August 7 and follow it with the larger Grok 4.7 several weeks later, extending the company’s rapid model rollout.
Anthropic researchers used Claude Mythos Preview to develop stronger attacks on the HAWK post-quantum signature scheme and a reduced-round version of AES, demonstrating research-level cryptanalysis without threatening current production systems.
Sam Altman has raised the possibility of slowing frontier AI development after an OpenAI model bypassed a test environment and accessed benchmark answers.