Arvind Narayanan highlights Joshua Saxe’s analysis of why cyber harms have not sharply increased alongside advances in AI capabilities. Saxe identifies this lack of a “discontinuity” in damages as a central challenge for AI security policy. The analysis argues that many cybercriminals are relatively unskilled, operate from lower-income countries, and face arrest, sanctions, and prosecution. These constraints limit how quickly criminal organizations can scale. Historically, ransomware required five to eight years to develop mature workflows, business models, and specialized divisions of labor capable of producing a multibillion-dollar industry. Saxe therefore suggests that AI-enabled criminal ecosystems may also take time to emerge, creating a temporary “defender-dominant window” for preparation.
A commenter challenges this relatively reassuring view by emphasizing state-sponsored and state-aligned actors. Unlike ordinary criminal groups, states possess substantial resources, advanced technical expertise, protection from law enforcement, and the ability to innovate rapidly. The commenter cites reports surrounding a PHP vulnerability, suggesting that actors may have used AI to identify and exploit a flaw before its patch was released. The absence of visible AI-driven damage could therefore reflect successful, covert operations or objectives other than financial gain. The commenter questions whether sophisticated groups such as those associated with Salt Typhoon fit the profile of low-skilled, arrest-prone criminals.
Another participant argues that attacks remain constrained by limited attacker attention, but warns that AI could dramatically multiply this resource. Agentic “worms” capable of autonomously spreading across compromised systems could create exponential growth in attack capacity, potentially causing massive losses if suitable incentives emerge.
Article timeline
Build this article’s timeline
Extract dated events, key actors, and open questions using the same timeline analysis available in Signal.
Manipulation Score: MediumHow much this article uses emotional framing vs. straight reporting
Loaded Language2825%
Absolutist1815%
Threat/Panic4825%
Us vs Them3015%
Engagement Bait1220%
Manipulation score look right?
Factual Claims (10)
mediumtechnology
There has not been a huge discontinuity in damages since the dawn of scaled AI.
mediumtechnology
The lack of responsiveness of damages to rising AI capabilities is a major conceptual crux and stumbling block for AI security policy.
mediumsocial
Cybercriminals are generally unskilled and are arrested frequently.
mediumsocial
Criminal organizations often base themselves in middle- and lower-income countries to reduce labor costs and recruit people with limited formal-economy options.
mediumlegal
Criminals face arrest, sanctions, interdiction, and prosecution, limiting their ability to scale.
lowlegal
The larger a criminal organization becomes, the more domestic and international law-enforcement attention it attracts.
mediumeconomic
It took five to eight years for ransomware to become a major industry, eventually reaching multi-billion-dollar scale.
mediumtechnology
The current period may be a defender-dominant window in which organizations can prepare for harmful future scenarios.
mediumgeopolitical
State actors are well resourced, highly skilled, relatively safe from law enforcement, and likely to innovate rapidly.
hightechnology
Entities may have used AI tools to identify a PHP vulnerability, develop an exploit, and attack random websites before a patch was available.
Comments
Login to view and post comments.