In early 2024, a hanging deepfake fraud case in Hong Kong introduced the vulnerabilities of AI-driven deception into sharp aid. A finance worker was duped throughout a video name by what gave the impression to be the CFO—however was, actually, a complicated AI-generated deepfake. Satisfied of the decision’s authenticity, the worker made 15 transfers totaling over $25 million to fraudulent financial institution accounts earlier than realizing it was a rip-off.
This incident exemplifies extra than simply technological trickery—it alerts how belief in what we see and listen to will be weaponized, particularly as AI turns into extra deeply built-in into enterprise instruments and workflows. From embedded LLMs in enterprise programs to autonomous brokers diagnosing and even repairing points in reside environments, AI is transitioning from novelty to necessity. But because it evolves, so too do the gaps in our conventional safety frameworks—designed for static, human-written code—revealing simply how unprepared we’re for programs that generate, adapt, and behave in unpredictable methods.
Past the CVE Mindset
Conventional safe coding practices revolve round recognized vulnerabilities and patch cycles. AI adjustments the equation. A line of code will be generated on the fly by a mannequin, formed by manipulated prompts or information—creating new, unpredictable classes of danger like immediate injection or emergent conduct outdoors conventional taxonomies.
A 2025 Veracode research discovered that 45% of all AI-generated code contained vulnerabilities, with frequent flaws like weak defenses towards XSS and log injection. (Some languages carried out extra poorly than others. Over 70% of AI-generated Java code had a safety difficulty, for example.) One other 2025 research confirmed that repeated refinement could make issues worse: After simply 5 iterations, important vulnerabilities rose by 37.6%.
To maintain tempo, frameworks just like the OWASP Prime 10 for LLMs have emerged, cataloging AI-specific dangers comparable to information leakage, mannequin denial of service, and immediate injection. They spotlight how present safety taxonomies fall quick—and why we’d like new approaches that mannequin AI risk surfaces, share incidents, and iteratively refine danger frameworks to mirror how code is created and influenced by AI.
Simpler for Adversaries
Maybe essentially the most alarming shift is how AI lowers the barrier to malicious exercise. What as soon as required deep technical experience can now be completed by anybody with a intelligent immediate: producing scripts, launching phishing campaigns, or manipulating fashions. AI doesn’t simply broaden the assault floor; it makes it simpler and cheaper for attackers to succeed with out ever writing code.
In 2025, researchers unveiled PromptLocker, the primary AI-powered ransomware. Although solely a proof of idea, it confirmed how theft and encryption might be automated with a neighborhood LLM at remarkably low price: about $0.70 per full assault utilizing business APIs—and primarily free with open supply fashions. That sort of affordability might make ransomware cheaper, sooner, and extra scalable than ever.
This democratization of offense means defenders should put together for assaults which are extra frequent, extra diversified, and extra artistic. The Adversarial ML Risk Matrix, based by Ram Shankar Siva Kumar throughout his time at Microsoft, helps by enumerating threats to machine studying and providing a structured strategy to anticipate these evolving dangers. (He’ll be discussing the issue of securing AI programs from adversaries at O’Reilly’s upcoming Safety Superstream.)
Silos and Ability Gaps
Builders, information scientists, and safety groups nonetheless work in silos, every with totally different incentives. Enterprise leaders push for fast AI adoption to remain aggressive, whereas safety leaders warn that shifting too quick dangers catastrophic flaws within the code itself.
These tensions are amplified by a widening abilities hole: Most builders lack coaching in AI safety, and plenty of safety professionals don’t totally perceive how LLMs work. Because of this, the outdated patchwork fixes really feel more and more insufficient when the fashions are writing and operating code on their very own.
The rise of “vibe coding”—counting on LLM options with out overview—captures this shift. It accelerates improvement however introduces hidden vulnerabilities, leaving each builders and defenders struggling to handle novel dangers.
From Avoidance to Resilience
AI adoption gained’t cease. The problem is shifting from avoidance to resilience. Frameworks like Databricks’ AI Danger Framework (DASF) and the NIST AI Danger Administration Framework present sensible steerage on embedding governance and safety instantly into AI pipelines, serving to organizations transfer past advert hoc defenses towards systematic resilience. The purpose isn’t to get rid of danger however to allow innovation whereas sustaining belief within the code AI helps produce.
Transparency and Accountability
Analysis exhibits AI-generated code is usually easier and extra repetitive, but in addition extra weak, with dangers like hardcoded credentials and path traversal exploits. With out observability instruments comparable to immediate logs, provenance monitoring, and audit trails, builders can’t guarantee reliability or accountability. In different phrases, AI-generated code is extra prone to introduce high-risk safety vulnerabilities.
AI’s opacity compounds the issue: A operate could seem to “work” but conceal vulnerabilities which are tough to hint or clarify. With out explainability and safeguards, autonomy rapidly turns into a recipe for insecure programs. Instruments like MITRE ATLAS might help by mapping adversarial techniques towards AI fashions, providing defenders a structured strategy to anticipate and counter threats.
Trying Forward
Securing code within the age of AI requires greater than patching—it means breaking silos, closing ability gaps, and embedding resilience into each stage of improvement. The dangers could really feel acquainted, however AI scales them dramatically. Frameworks like Databricks’ AI Danger Framework (DASF) and the NIST AI Danger Administration Framework present constructions for governance and transparency, whereas MITRE ATLAS maps adversarial techniques and real-world assault case research, giving defenders a structured strategy to anticipate and mitigate threats to AI programs.
The alternatives we make now will decide whether or not AI turns into a trusted accomplice—or a shortcut that leaves us uncovered.
