"Offensive AI is the moment the weapon learns to aim itself." — Foster Nethercott
Script kiddie as a service: the shifting attacker profile
Foster Nethercott argues that agentic AI has erased the old skill barrier. Where once an attacker needed technical mastery to write exploits or craft convincing social engineering, the author says "intent and access to capable tools suffice." He coins the term "script kiddie as a service" to describe how previously unskilled actors can now launch sophisticated, autonomous attacks by leaning on agents that plan and execute campaigns without a human pulling the trigger.
That change narrows the constraints on attackers from human ability to the capabilities of the AI model selected. Nethercott cautions that as many untrained actors reuse the same models and methods, their operations converge into a "behavioral monoculture" — predictable, high-volume chains of standardized phishing and exploit patterns. While skilled adversaries will diverge from those defaults, the bulk of new attackers may not, creating both a surge in attacks and detectable commonality in their techniques.
Autonomous social engineering: the hunt that runs itself
Nethercott lays out a concrete example of an agentic workflow: one agent collects open-source intelligence (LinkedIn profiles, press releases, conference recordings) to build a target profile; a second agent composes, sends and manages personalized messages; the conversation advances without further human intervention. The result, he writes, is not merely faster attacks but the erosion of the defensive cues defenders historically relied on — "the clumsy grammar, the recycled template, the identical mail sent ten thousand times."
As a consequence, defenders must shift reliance toward infrastructural signals — sender reputation, authentication, and similar controls — because the linguistic and template-level tells are being removed by fluent, individualized messages generated by agents.
Exploitation, malware, and the false confidence of retrieval
Nethercott traces the same agentic automation into vulnerability discovery and malware development. He notes that models trained to chain tool calls and self-correct make exploit generation easier, and that integrating even moderately capable models with a retrieval database of known vulnerabilities lets an agent "perform its own reconnaissance" and select likely exploits from a shelf.
Two specific examples in the piece underline the scope of concern: Nethercott references his earlier introduction of the "Guided Network Access Weapon (GNAW)" at the Hackers Teaching Hackers conference as an early sign of this trajectory, and he states that "the federal government is now getting involved and forcing models like Anthropic's Fable 5 to be taken off the market over fears of its capabilities." He warns that retrieval‑based outputs can be plausibly related rather than genuinely applicable — an agent "does not check the version, nor the configuration, nor whether the service can even be reached" — producing confident but potentially hollow conclusions.
SANS Secure AI Blueprint and the primacy of Utilize
Nethercott cites the SANS Secure AI Blueprint, authored by SANS Chief AI Officer Rob T. Lee, which divides AI security into three tracks: Protect AI, Utilize AI, and Govern AI. He emphasizes that the Utilize track — the operational use of AI for offense and defense — is where proofs are made. "A defense is a theory until it makes contact," he writes, arguing that offensive testing against real systems reveals whether governance and hardening hold under attack. In Nethercott's view, operators who run these exercises increasingly "hold the whole program to account."
What this means for operators, policymakers, and enterprises
- Operators and security teams: Nethercott urges hands-on experience. His SEC535: Offensive AI – Attack Tools and Techniques course at SANS San Antonio 2026 (three days, hands-on labs) is presented as a way for operators to "drive the tooling with your own hands" and learn where agents can be trusted and where human judgment must intervene.
- Policymakers and regulators: The article notes federal involvement in model availability (the example of Fable 5) and highlights the role of Governance in the SANS framework, implying that oversight and policy decisions will shape how powerful models are distributed and used.
- Affected enterprises and procurement leaders: Nethercott's point that defenders' assumptions about linguistic tells no longer hold means organizations must prioritize infrastructure-level controls and validated red‑team exercises that use modern agentic tooling to test defenses in realistic conditions.
Nethercott's core caution is sharp and final: the machine can now aim itself, but it cannot decide whether the shot should be taken. He frames the remaining human role as judgment — the single responsibility that has grown to encompass nearly the whole of the work. "The weapon no longer needs a warrior to swing it," he writes, "but it has never needed a person to decide whether it should be swung at all more than now."
Read the original piece: https://thehackernews.com/2026/06/agentic-ai-weapon-that-no-longer-needs.html
