Artificial intelligence is increasingly discussed in information security contexts, not only as a tool for threat detection or development automation but also for vulnerability discovery, attack preparation, and exploitation of software weaknesses.
One common prediction is that continued AI progress will inevitably drive a rise in cyberattacks. The reasoning appears straightforward: tasks that once required specialized knowledge, time, and effort can now be partially handled by large language models and autonomous agents. When several models are combined with tooling and granted the ability to plan actions independently, human involvement could theoretically be limited to task definition.
Yet these assumptions require closer scrutiny. The fact that AI can identify a vulnerability does not automatically mean an unprepared individual can exploit it. Likewise, successful performance on laboratory tasks does not prove an agent can independently compromise real-world systems.
AI also lowers barriers to software creation. Individuals without engineering backgrounds can now build web applications, connect databases, implement authentication, and deploy services online. However, the ability to produce working code does not imply understanding of architecture, limitations, or security requirements.
Defenders are adopting AI at a comparable pace. Automated source-code analysis, bug detection, and patch generation are advancing as rapidly as offensive tooling.
PHP2Shell: AI-assisted discovery of complex vulnerabilities
On 21 October 2026, the PHP project released security updates addressing two critical core interpreter vulnerabilities: CVE-2026-109618 and CVE-2026-109619, collectively referred to as PHP2Shell. Both issues resided in low-level memory management and error-handling mechanisms and could enable remote code execution, including unauthenticated scenarios.
The first flaw, CVE-2026-109618, stemmed from incorrect state handling during memory exhaustion, leaving partially initialized internal structures that were later treated as valid. The second, CVE-2026-109619, involved retention of incorrect global state after a fatal error, creating use-after-free conditions.
Initial indicators of CVE-2026-109618 had been noted during a 2024 external security audit conducted by Quarkslab and OSTIF. At that time, no practical exploitation path to remote code execution was identified, and remediation efforts were complicated by compatibility and performance concerns. In 2026, teams including Calif, OpenAI, and Wiz used AI-assisted analysis tools such as Claude Mythos and OpenAI Daybreak Red to demonstrate practical exploitability.
The PHP Foundation emphasized that AI did not discover the issues from scratch; human researchers had already identified suspicious behavior years earlier. AI contributed by helping analyze codebases, generate hypotheses, and validate exploitation scenarios.
RAND study on AI uplift for offensive operations
On 28 May 2026, the RAND Corporation published “Investigating the potential use of frontier AI models for offensive cyberattacks: A human uplift study,” commissioned by the UK AI Security Institute. The randomized controlled trial involved 157 participants with varying technical backgrounds and tested tasks covering network operations, operating-system exploitation, and vulnerability chaining.
Participants using models such as OpenAI o3, GPT-5, Claude Opus 4.1, Claude Sonnet 3.7, and Gemini 2.5 Pro showed an 18-percentage-point increase in success on the simplest scenario among novices, compared with a 5-percentage-point gain among experienced participants. No statistically significant advantage appeared for completing full attack chains. The study noted an “onboarding uplift” effect that helped beginners understand initial context but did not reliably translate into end-to-end success.
Researchers concluded that while AI lowers entry barriers for certain subtasks, it does not eliminate the requirement for technical judgment when adapting suggestions to differing environments or recognizing incorrect outputs.
