Don’t be fooled by this summer of AI hype

Anthropic said on April 30 that its new Claude Mythos model can spot software vulnerabilities better than most human security experts, a claim that sparked a wave of media attention and industry debate. The announcement came just weeks after a high‑profile breach in which OpenAI’s GPT‑4 and Hugging Face’s open‑source models were used to extract proprietary code from a corporate network, prompting both companies to tighten access controls.

The hype reflects a broader shift toward AI‑driven security tools, a market that has grown sharply as cyber‑attacks become more sophisticated. Companies are racing to embed large language models into their defenses, hoping the models’ pattern‑recognition abilities can keep pace with evolving threats. However, the rush has also exposed weaknesses: the OpenAI–Hugging Face incident showed that even well‑guarded APIs can be coaxed into revealing sensitive information, and it raised questions about the reliability of model‑generated security advice.

Anthropic and Meta later disclosed similar incidents in which their models inadvertently leaked internal data during testing. Anthropic described the leak as “proudly” reported, emphasizing its commitment to transparency, while Meta framed the episode as an “unintended” exposure that it is investigating. Security experts warned that these disclosures highlight a systemic risk: large language models can memorize and regurgitate proprietary code or confidential text, especially when trained on massive, uncurated datasets.

Industry observers say the next step will be stricter oversight and better model‑training practices. Regulators in the EU and the U.S. are drafting rules that could require AI providers to certify that their systems do not expose sensitive data. Meanwhile, firms are investing in “red‑team” exercises to probe their own models for leaks before deployment. The summer of AI hype may be winding down, but the push for safer, more accountable AI in cybersecurity is only just beginning.