A wave of high-profile announcements from leading artificial intelligence companies this summer has been met with skepticism from independent experts, who argue that many of the reported breakthroughs are overstated marketing rather than scientific milestones. Recent claims by Anthropic and OpenAI regarding security vulnerabilities, mathematical proofs, and rogue AI agents have dominated headlines, but subsequent expert analysis suggests a different reality.
What Happened
In late April, Anthropic claimed its model Claude Mythos was superior to most security experts at identifying software vulnerabilities. This was followed by an incident involving OpenAI and Hugging Face, which Anthropic and Meta described as hacking events caused by their models. Anthropic also claimed a mathematical breakthrough, a feat OpenAI soon reported with its own model, Astra. Additionally, Anthropic engineer Jacob Coxon recently departed the company, stating that Anthropic and OpenAI are "racing straight towards self-improving superintelligence and gambling with our lives."
However, cybersecurity experts have countered the narrative of "models gone rogue," attributing the hacking incidents to OpenAI’s negligence and failure to implement basic security practices rather than autonomous AI behavior. Regarding the mathematical claims, initial excitement among mathematicians about Astra solving long-standing problems faded upon closer inspection. Experts concluded the results were not as novel as initially presented, with some accusing OpenAI of research misconduct and plagiarism. Tristan Buckmaster, a professor at NYU’s Courant Institute, published a statement suggesting OpenAI had improperly attributed work from other researchers.
Why It Matters
The discrepancy between corporate press releases and expert analysis highlights a growing tension in the AI industry. The reliance on fields like computer programming and mathematics for demonstrating AI capabilities is strategic; these domains allow for automated verification of outputs, reducing the need for expensive human annotation. However, this has led to concerns that corporations are exploiting the prestige of these fields to exaggerate product capabilities. A statement signed by hundreds of mathematicians warned of a "strong commercial incentive" to overstate AI performance and urged policymakers to rely on expert consultation rather than press releases.
Furthermore, the framing of AI products as "rogue" or "superintelligent" shifts accountability away from the companies developing them. By describing software as having agency, firms can deflect criticism regarding data privacy, plagiarism, or security failures. This narrative also influences policy, as seen in legislative proposals aimed at preventing "artificial superintelligence," which critics argue are based on ideological narratives rather than current engineering realities.
The Bottom Line
While AI companies continue to announce significant advancements, independent experts argue that the current hype cycle obscures fundamental issues such as corporate negligence and research integrity. The industry's tendency to anthropomorphize models serves to market products as incipient AGI, potentially misleading the public and policymakers about the actual state of the technology.