AI leaders and experts are gathering on Duke University's campus in Durham on Friday to address growing concerns about dangerous artificial intelligence, with the 9:15 a.m. event kicking off the third annual Cybersecurity at Duke conference. The gathering is not open to the general public, but its subject matter has become impossible to ignore: whether the technology racing through American industry is also racing beyond anyone's ability to control it.
A Researcher's Exit Sets the Stage
The timing lands just over a week after Jacob Coxon, a former researcher for Claude AI who had previously worked at OpenAI before joining Anthropic, announced he was resigning from the company over AI safety concerns, according to Forbes. Claude AI is the AI assistant built by Anthropic, and Coxon estimated a 10% chance of AI causing human extinction within the next decade. He argued that AI companies, in his words, “are racing straight to self-improving superintelligence and gambling with our lives,” as reported by Fox Business.
Coxon was not alone in that number. Evan Hubinger, Anthropic's alignment science lead, said he personally believed there was a greater than 10% chance AI could kill all humans within the next decade, telling the BBC, “I personally think it is >10% within the next decade.” Hubinger drew a distinction, however, saying the danger from today's chatbots and tools is not the concern. “The risk from present models is low,” he said, per Fox Business, with the real worry centered on a future superintelligence emerging through recursive self-improvement. Hubinger also said Anthropic does not yet have a plan to solve the issue of alignment for superintelligence and is not clearly on track to do so.
Slowdown Calls Meet Skepticism
Anthropic CEO Dario Amodei has called for the AI industry to slow development, describing the risks as serious, according to the BBC. That call fits into a broader push: several leading AI researchers and executives signed a statement titled “Pacing the Frontier” in July, calling for the option to slow frontier AI development to address emerging risks and strengthen oversight, Forbes reported. OpenAI chief scientist Jakub Pachocki echoed the sentiment, telling CNBC, “I expect and hope for voluntary slowdowns to become commonplace until shared safety bars are established.”
Not everyone in the industry agrees the alarm is warranted. Some industry figures have suggested the extinction warnings are overstated or designed to generate hype, the BBC noted, and those warnings have grown markedly more stark in recent weeks as companies struggle to keep their own systems in check.
Who's Speaking at Duke
Friday's program in Durham will feature a discussion about the rapid deployment of autonomous AI systems across enterprise security operations and the questions that deployment raises, according to WRAL. Tim Kosiba, the Trump administration-appointed National Security Agency deputy director, is scheduled to speak and will focus on how adversaries could weaponize AI. Democratic Congresswomen Valerie Foushee and Deborah Ross are also slated to appear, and the pair will discuss the role legislation could play in safeguarding cybersecurity systems against AI, per WRAL's report.
The lineup reflects a split in Washington's posture toward the technology. President Donald Trump has touted AI as an important economic engine even as he has called fears around AI a hoax, WRAL reported — a tension that underscores why lawmakers like Foushee and Ross are expected to press the legislative angle at Friday's session.
The Cybersecurity Stakes Behind the Debate
The philosophical debate over extinction-level risk has a more immediate cousin: everyday security failures. OpenAI, Anthropic and Meta all disclosed hacks carried out by their AI tools over the summer, according to the BBC. Separately, cybersecurity researchers at SentinelOne have cataloged a broader set of risks built into modern AI systems, including data poisoning, adversarial attacks, model theft, privacy leakage, and API or hardware exploits. Data poisoning in particular can corrupt an AI system's training data in ways that modify its functionality and produce false choices or predictions, per SentinelOne.
Mitigating those risks, SentinelOne notes, requires validated data pipelines, strong access and API controls, adversarial testing, continuous monitoring, and incident-response plans — the kind of practical, unglamorous defenses that sit alongside the loftier debate over superintelligence. For Durham, a city increasingly shaped by Duke's research footprint and the region's tech industry, Friday's conference puts both ends of that spectrum, from existential risk to routine data breaches, on the same stage.
