Bright Simons, Vice President of IMANI Africa and one of the continent’s most prominent technology policy voices, has warned that widespread public misreading of how artificial intelligence (AI) systems become more capable is creating dangerous conditions for policymaking, using the global reaction to Anthropic’s Claude Mythos model as his primary evidence.
Simons argued that the public debate following Mythos’ unveiling conflated two fundamentally different things: improvements in machine reasoning, and the expansion of AI systems’ ability to autonomously operate across large networks of human-built digital tools. Treating these as the same thing, he said, produces both inflated fears and misguided policy responses.
Anthropic released a preview of Mythos in April 2026 under Project Glasswing, granting a small group of major technology companies including Amazon, Apple and Microsoft early access to the model specifically for defensive cybersecurity work. The company said the model had identified thousands of zero-day vulnerabilities, including some in every major operating system and web browser, with one bug having gone undetected for 27 years.
The findings prompted such alarm that Bank of England Governor Andrew Bailey, also a member of the Financial Stability Board (FSB), requested that Anthropic brief financial regulators on the cybersecurity implications, with FSB members growing concerned about exposure in banks’ digital defenses.
Simons said the public concluded from these results that AI had made a sudden leap toward superhuman intelligence. He disagreed. “AI has suddenly grown super-smart” is the wrong interpretation, he argued, saying that benchmark comparisons with rival frontier models showed relatively modest gains in areas tied to human-style cognition such as search, knowledge retrieval and browser operation. What Mythos demonstrated instead, he said, was strength in autonomously combining and manipulating large numbers of external tools across long-horizon tasks in ways difficult for humans to anticipate.
“That is right. We are giving AI tools autonomous control of a vast range of human-created tools,” Simons wrote, framing this as the genuine source of the model’s power and the actual locus of risk.
His most striking observation concerned cybersecurity infrastructure. He argued that virtually all existing software security architecture was built around the assumption that attackers would be human, operating at human speed, with human limitations. Agentic AI systems capable of running continuously at machine scale across multiple platforms simultaneously render much of that architecture obsolete. “Almost all of our software security was built in the belief that it is a human that will be trying to break in,” he said.
That concern is backed by data. AI-enabled cyberattacks increased by 89 percent in 2025 compared with the prior year, according to cybersecurity firm CrowdStrike, with AI agents proving more effective than most humans at locating and exploiting software vulnerabilities.
To make the risk intuitive, Simons compared giving an advanced AI system access to digital infrastructure to placing “a highly energetic child” in control of missiles, satellites and drones through a gaming console. The child’s energy and creativity become dangerous not because of intelligence but because of the reach of the tools placed in its hands.
Simons said the confusion stems from a fundamental failure to understand that advanced AI systems “harness human intelligence” rather than generating independent omniscience. He announced plans to promote what he calls the Social Edge Framework, an initiative aimed at examining the policy and business implications of increasingly autonomous AI agents for governments and institutions.
He warned that failure to correct this conceptual gap would lead not only to poor regulation but to dangerous institutional assumptions about what AI can and cannot do on its own.
