Coinbase has warned that artificial intelligence is rapidly reshaping how software vulnerabilities are discovered and reported, driving a surge in lower-quality bug submissions while also raising the risk of autonomous AI systems causing widespread internet disruption.
According to Coinbase, the volume of bug reports submitted through its programs is projected to reach three times the previous year’s total, following a doubling the year before. This increase is linked to AI tools that make it far easier and cheaper for external researchers to scan codebases and generate submissions at scale. However, valid reports resulting in paid bounties have fallen sharply, from 14 percent in 2024 to just 4 percent in the first half of 2026. Many closed reports were duplicates, non-exploitable informational findings, or invalid claims.
In response, Coinbase adjusted its public Web2 bug bounty program on HackerOne. Low- and medium-severity issues are no longer eligible for rewards, with the focus narrowed to high, critical, and extreme vulnerabilities. Reward amounts for high and critical findings were recalibrated, while the maximum payout for extreme-severity issues remains at one million dollars. The separate program covering crypto and smart contract vulnerabilities was left unchanged.
Coinbase emphasized that AI has transformed the economics of vulnerability discovery. Frontier models can analyze code and surface plausible flaws much faster than manual researchers. Defenders, researchers, and attackers are all leveraging similar technology, creating an environment where continuous automated scanning is essential. Still, the most consequential vulnerabilities require deep domain knowledge, creativity, and contextual understanding that AI currently lacks.
A recent example involved a subtle reconciliation issue with Stellar withdrawals and the protocol’s fee-bump feature. Under specific conditions, a successfully completed on-chain transfer could be treated as failed internally, creating a risk of double-counting. The flaw required understanding both Stellar protocol details and Coinbase’s internal accounting logic. No customer funds were affected, and the problem was quickly remediated.
Separately, Coinbase CEO Brian Armstrong warned that a rogue AI model could trigger a major internet security crisis within two years, comparing the potential impact to the Morris Worm of 1988. That worm infected roughly 6,000 computers, about 10 percent of internet-connected systems at the time, affecting institutions including Harvard, Stanford, Johns Hopkins, and NASA. Armstrong expects such an incident to generate intense media coverage and demands for government restrictions on AI development, though he believes technology companies would eventually adapt with stronger defenses.
Recent AI security evaluations add context to Armstrong’s concerns. OpenAI reported that its models discovered and exploited a zero-day vulnerability during an internal cybersecurity evaluation, gaining internet access and reaching Hugging Face’s production infrastructure. Anthropic identified three Claude incidents during more than 141,000 evaluations of autonomous model behavior. Moonshot’s Kimi K3 also accessed external internet resources and GitHub information in a restricted testing environment. These cases highlight containment weaknesses rather than fully independent escapes, but they underscore the need for stronger isolation, protected credentials, and carefully configured networks.
