Vibe Coding Security: Game Jams Hide a Growing Flaw Debt

From meme to game jam

Vibe coding, the practice of describing what you want and letting an AI model write the software, has moved well beyond its origins as an online catchphrase. Recent signals include public resource lists for "vibecoders," a dedicated AI and vibecode community event in Prague, job postings seeking AI-native production skills, and a game jam built around vibecode.game 4. That last item is arguably the most telling. Games require state management, logic, pacing, assets and interaction design, which is far more than a landing page, so their arrival suggests people now expect AI to carry ideas all the way to working products 4.

That cultural momentum is real. But a cluster of security research published over the past several months points to a less flattering picture. The code being shipped this way is frequently insecure, and much of it is already running in production.

The numbers keep converging

The most striking comparison comes from a three-month study by Symbiotic Security, published August 27. It covered 1,967 public GitHub repositories and 1,072 live vibe-coded apps across five platforms, cross-checked with ten independent scanners, and turned up 87,826 vulnerabilities in total 2. AI-assisted repositories averaged 42.3 vulnerabilities each, compared with 9.6 for human-only repositories, roughly a 4.4x gap. Some 98% of the vibe-coded apps had at least one flaw 2.

Academic work lands in the same place. Researchers behind "Understanding the (In)Security of Vibe-Coded Applications" assembled a corpus of 9,041 open-source vibe-coded apps, then randomly sampled 200 publicly accessible deployments for auditing with a workflow that combined multi-agent code review and human validation 3. They found that 91.0% contained at least one vulnerability, and 65.77% of the identified issues were rated Critical or High. The flaws clustered in broken access control, injection and authentication failures 3. A secondary summary of the same study reports 180 of 200 apps (90%) vulnerable and 1,471 vulnerabilities, with broken access control making up 36% 2. The one-point gap between 90% and 91% looks like a rounding or revision difference rather than a substantive disagreement.

Benchmark results add to the concern. The SUSVIBES study, revised in August, ran 186 real-world tasks across 12 agent configurations. The best setup produced functionally correct code 57% of the time but secure code only 11.8% of the time, and 79.3% of the correct solutions were still vulnerable 2. Put plainly, "it works" and "it's safe" are separate outcomes, and current agents are much better at the first.

Why the flaws keep appearing

The arXiv authors attribute the vulnerabilities to eight recurring failure modes rooted in three systematic limitations of AI agents: memory defects, objective defects and knowledge defects 3. That framing matters because it describes structural weaknesses in how agents approach a task, not occasional bad luck. An agent that loses track of earlier context, optimizes narrowly for a working feature, or lacks current security knowledge will reproduce the same mistakes at scale.

The visible part is probably small

Georgia Tech's School of Cybersecurity and Privacy is approaching the problem from the disclosure side. Its Vibe Security Radar scanned more than 43,000 security advisories and has so far confirmed 74 cases tied to AI-generated code. Fourteen were critical and 25 high, including command injection, authentication bypass and server-side request forgery 1. Graduate researcher Hanqing Zhao noted that nobody had been tracking such vulnerabilities before the tool launched 1.

The researchers themselves say they have only scratched the surface. The radar identifies AI involvement through metadata such as co-author tags, bot email addresses and known tool signatures, and it cannot attribute a flaw once those markers are stripped 1. Seventy-four confirmed cases should therefore be read as a floor, not an estimate.

Reading the moment

These sources describe two trends moving in opposite directions. Adoption is broadening into richer, more stateful software, and the people building it increasingly treat vibe coding as a craft and a job skill 4. At the same time, independent studies using different methods (repository scans, live-app audits, agent benchmarks and advisory mining) all find insecurity to be the default rather than the exception 123.

In my reading, the game-jam milestone is less a sign of maturity than a sign of reach. Ambitious projects mean more state, more user data and more access-control logic, and those are the areas where the audits find the worst failures. Until security review becomes a standard step in the prompt-to-production pipeline, each new category vibe coding enters is likely to add to a security debt that someone will eventually have to pay.