2026 GenAI Code Security Report: AI Is Writing More of Your Code but Security Hasn’t Caught Up
New GenAI code security research shows a stubborn truth: as AI is generating more of the code entering production, secure output is not improving at the same pace. The result is a GenAI code security challenge defined by scale, model choice, and the growing need for verification.
AI coding has moved past experimentation. For many teams, it is now part of how software gets built every day.
That’s the opportunity. It’s also the risk.
In organizations that have adopted AI coding tools, AI now authors roughly half of all committed code. That changes the GenAI code security conversation completely. This is no longer about whether AI can write code. It can, and it does, at production scale. The real question is whether security programs can keep up with the volume, speed, and variability of what those systems produce.
Our latest research shows the answer is still no.
The 2026 GenAI Code Security Report found that roughly 44% of AI code generation tasks introduced a risky security vulnerability in tests. The average security pass rate across models is 56% – barely changed from 55% in the first report. In other words, security performance has stayed flat while the amount of AI-generated code entering pipelines has surged.
That is why GenAI code security is now a scale problem, not a theoretical risk discussion. If AI is responsible for half the codebase and vulnerable output remains this common, every organization needs a sharper strategy for model selection, verification, remediation, and governance.
AI code volume is rising faster than control mechanisms
The biggest shift in this year’s report is not a dramatic collapse in security performance. It is that the same failure rate now applies to much more code.
Security teams are already feeling the pressure. As AI speeds up software creation, code review expands, remediation queues grow, and governance models strain to keep up with workflows that are changing in real time. Leaders are not looking for a way to shut AI down. They want a way to scale it safely, with confidence.
That is the crux of the market moment. AI is increasing developer throughput, but it is also increasing the number of security decisions being made upstream, often before security teams have visibility or context. When verification cannot keep pace, risk accumulates quietly.
This is why GenAI code security cannot be treated as a side discussion inside a broader AI strategy. It is becoming a core part of software trust.
The headline finding: syntax is solved, security is not
The report data makes one point unmistakably clear: functional output and secure output are fully separate.
Modern models now generate syntactically correct code nearly 100% of the time. Syntax is effectively solved. But secure coding is not following the same curve. Across all languages, CWEs, tasks, and models tested, the average security pass rate remains at 56%.
That gap matters because syntactic fluency can create false confidence. Code that compiles, runs, and looks clean can still introduce exploitable weaknesses into production. In fact, roughly 44% of tested AI code generation tasks resulted in code with a known vulnerability.
For leaders making platform and policy decisions, this is a critical distinction. The ability to generate working code faster is valuable. But working code is not trustworthy code by default. GenAI code security depends on verification, not appearances.
Why this matters now
This research is landing at a time when software risk is already intensifying.
The 2026 Verizon Data Breach Investigations Report found that software vulnerabilities are now the top breach entry point at 31%, surpassing stolen credentials. At the same time, Veracode’s 2026 State of Software Security shows that security debt now affects 82% of organizations, critical security debt affects 60%, and high-risk vulnerabilities are up 36% year over year.
Put those signals together and the implication is clear: this is not just about more AI-generated code. It is about more vulnerable code entering environments that are already carrying too much unresolved risk.
For CISOs and AppSec leaders, that changes the operational question. The issue is no longer whether AI-generated code introduces risk. The issue is how to prevent that risk from overwhelming review capacity, expanding remediation backlogs, and weakening governance at enterprise scale.
This is also why safe enablement matters more than blanket restriction. The organizations moving forward successfully are not trying to slow development to a crawl. They are trying to build a system where AI-assisted development can move fast while security teams maintain visibility, context, and control.
What the data says about model choice
One of the most important findings in the report is that model selection has become a security decision.
The average may be flat, but the field is not uniform. GPT-5.5 leads the current snapshot at 68%, while more than half of tested models sit at 50–53%. That spread has practical consequences. A model that fails one in three security tasks creates meaningfully less downstream remediation work than one that fails one in two, even if that remediation is AI-assisted.
For enterprises scaling AI coding, that difference matters. It affects how many flaws need to be found, how much developer time is spent fixing them, and how much burden lands on AppSec and engineering leaders to retest and govern the output.
The takeaway is simple: not all productivity gains are equal. If two models help developers move fast, but one consistently produces less vulnerable code, that is not just a technical preference. It is a risk management decision.
The best-performing model still does not solve the problem
GPT-5.5 leads at 68%. That is notable progress relative to the rest of the field.
It also means the top model still fails on nearly one in three security tasks.
That is the right way to read the leaderboard. A handful of models are pulling ahead, but no model in this dataset generates secure code reliably enough to remove the need for verification. Better model choice can reduce load. It does not eliminate risk.
More capability does not automatically mean more security
Some of the most useful findings in this year’s report challenge common assumptions.
Coding-specialized models averaged 51% on security tasks. General-purpose models averaged 52%. Models purpose-built to write code faster are not, on average, writing it more securely.
Reasoning models do show an edge, averaging 56% versus 51% for non-reasoning models. That suggests internal reasoning may help models catch insecure constructs before output. But even the stronger category still fails on nearly 44% of tasks.
Model size also does not change the picture much. Large models averaged 53%, while medium and small models each averaged 51%. Bigger models are not solving the secure coding problem either.
Together, these results point to a broader reality: security is not a side effect of general model progress. It requires deliberate emphasis, careful evaluation, and strong controls around how models are used in production workflows.
Where vulnerabilities persist tells you where verification matters most
The report also found that performance varies dramatically by vulnerability type.
On SQL injection, models performed relatively well, with an average pass rate of 83%. On cryptographic algorithms, they reached 87%. But performance fell sharply on cross-site scripting at 15% and log injection at 12%.
That spread matters because it reveals where teams can and cannot trust default output.
Some security problems are easier for models to learn as repeatable patterns. Others depend on dataflow, application context, and a deeper understanding of how user input moves through systems. Those are the areas where raw model output is most likely to fail and where verification needs to be strongest.
For security leaders, this is the practical lesson: GenAI code security is not one uniform control problem. It is a prioritization problem. Teams need to know which kinds of weaknesses remain persistent, where context is missing, and how to focus reviews where models are least dependable.
This is what scaling safely looks like
The organizations that will benefit most from AI coding will not be the ones that assume the model is safe. They will be the ones that operationalize trust.
That means treating GenAI code security as a discipline built on four capabilities:
1. Choose models with security in mind
Performance differences across models are meaningful. Procurement and platform decisions should include security evaluation, not just cost, speed, or developer preference.
2. Verify code in the workflow
Because syntax is no longer the issue, teams need controls that validate security as code is written and before it reaches production. The faster AI writes, the more important in-workflow verification becomes.
3. Focus remediation where risk is highest
When vulnerable output increases, backlog management becomes a strategic issue. Security teams need unified visibility to prioritize exploitable, high-impact flaws and reduce the noise that slows action.
4. Govern AI coding like a real software input
If AI is writing half the code, AI usage needs policy, accountability, and measurable oversight. Governance has to cover what models are used, where they are used, how output is checked, and how results are tracked over time.
This is where a unified approach matters. Teams need to find risk early, fix it quickly, and govern it consistently across the software lifecycle. That is how secure software delivery scales in the AI-coding era.
The next phase of GenAI code security is operational
The conversation around AI coding is changing. The question is no longer whether the technology is impressive. It is whether organizations can absorb its output without expanding risk faster than they can reduce it.
Our research shows the answer depends on what happens next.
When AI authors roughly half of committed code, a flat 56% security pass rate is not acceptable as background noise. It is an operational signal. It means model choice matters. Verification matters. Prioritization matters. Governance matters.
Most of all, it means secure software delivery needs to evolve as quickly as AI-assisted development itself.
Read the full 2026 GenAI Code Security Report to see which model choices reduce risk, where vulnerabilities persist, and how to strengthen GenAI code security as AI coding scales.