His message? Tools using AI to help write code could actually open doors to fresh risks on user devices.
Not everything about coding assistants runs smoothly, Vanunu pointed out during his talk. Tools like Claude Code, OpenAI Codex, and Google Gemini carry hidden flaws despite their popularity. Though they speed up work for programmers, deeper issues emerge beneath the surface. Security measures that have stood firm for years now face quiet circumvention.
What looks like progress might also open backdoors by design.
Despite gains in digital protection during recent years - tools like real-time threat tracking, isolated testing environments, and internet-hosted setups have made devices safer - an unforeseen setback is emerging. Artificial intelligence helpers used in software creation now demand broad entry into internal machines, setup records, along with connection points. Since coders routinely allow full control, unseen doors open.
These openings can be used by hostile actors aiming to infiltrate. Progress, it turns out, sometimes carries hidden trade-offs.
Now under pressure from AI agents wielding elevated access, Vanunu likened today’s endpoints to a once-solid fortress. These tools, automating actions while interfacing deeply with system settings, slip past conventional defenses unable to track such dynamic activity.
A blind spot forms - silent, unnoticed - where malicious actors quietly move in.
One key issue identified in the study involves the exploitation of config files like .json, .env, or .toml. While not seen as harmful by many, such file types typically escape scrutiny during security checks. Hidden within them, hostile code might reside - quietly waiting. Because systems frequently treat these documents as safe, automated processes, including AI-driven ones, could run embedded commands without raising alarms.
This opens a path for intrusion that skips conventional virus-like components.
Unexpected weaknesses emerged within AI coding systems, revealing gaps like flawed command handling. Some platforms allowed unauthorized operations by sidestepping permission checks. Running dangerous instructions became possible without clear user agreement in certain scenarios. Previously accepted tasks were altered silently, inserting harmful elements later. Remote activation of external code exposed further exposure points.
Approval processes failed under manipulated inputs during testing.
Even after fixing these flaws, one truth stands clear - security boundaries keep changing because of artificial intelligence. Tools meant to help coders do their jobs now open new doors for those aiming to break in. What once focused on systems has moved toward everyday software assistants. Fixing old problems does not stop newer risks from emerging through trusted workflows.
Starting fresh each time matters when checking every AI tool currently running. One way forward involves separating code helpers into locked-down spaces where they can’t reach sensitive systems. Configuration files deserve just as much attention as programs that run directly. With more companies using artificial intelligence, old-style defenses might no longer fit the real dangers appearing now.
