Two reports landed a day apart this week, and they make more sense read together. Veracode’s latest GenAI Code Security Report says the code coming out of large language models is getting safer for the first time since the series began in July 2025. SecurityScorecard’s healthcare study, built on its Project Meridian mapping platform, shows the other end of the pipeline: the internet-facing systems of almost 40,000 US hospital and clinic sites, where known flaws sit waiting. One report is about how vulnerabilities get written. The other is about how long they stay.
AI code got better, and 38 percent still fails
Veracode gave 18 models coding tasks in Python, C#, JavaScript and Java, without asking them to write securely, and checked the output for four weakness types: SQL injection, weak cryptography, cross-site scripting and log injection. The average security pass rate rose from 56% to 62%. Turned around, 38% of tasks still produced code with a known vulnerability, down from 44%.
The spread between models is wide. Claude Opus 5.5 and GPT-6-Sol tied at the top with 73%, while the weakest models still score in the mid-40s. Ten models now clear 60%, against four in the previous round. Language matters as much as the model. Python passed about 64% of the time, C# about 60% and JavaScript about 57%. Java sits near 31%, so roughly two out of three Java tasks failed.
By weakness type the old pattern holds. Models handle SQL injection and crypto choices fairly well. Cross-site scripting improved the most, from a very low base, and log injection is still stuck around 12%. Those are the bugs that depend on where data goes after the function returns, and models still don’t follow it there.
Veracode’s conclusion is that a mid-tier model plus automated scanning beats a frontier model with no checks. It sells scanning, so take that with salt. The arithmetic still works against defenders. If a team doubles the code it ships with AI help and the flaw rate per task falls from 44% to 38%, more flaws go out the door than before.
What attackers see at the hospital door
SecurityScorecard came at the problem from outside. Its report, titled Visiting Hours, uses only what any attacker can see from the internet to assess 39,319 healthcare records. Each record is a mapped site, not a whole health system.
About one in seven records, 5,862 in all, carries a flaw from CISA’s Known Exploited Vulnerabilities catalog, meaning attackers already use it in real attacks. Of the 7,818 records that could be graded, 62.5% earned an A or B, and one in six earned a D or F. Leaked-credential signals showed up on 3,767 records, about 9.6%, and 10 records had signs consistent with possible compromise.
A good grade can hide a lot. One Texas health system earned an A and still had eight urgent vulnerabilities on live machines, the most of any group studied. In the same Texas sample, 11 of 12 hospital groups relied on the same vendor, and 72 of 313 suppliers turned up at two or more groups. The 2024 Change Healthcare attack showed what one shared supplier can do to pharmacies and billing across the whole country.
Rural clinics get a D or F slightly more often than hospitals, 16.7% against 15.2%. The gap in grades is small. The gap in consequences is not. When a city hospital goes down, ambulances divert a few blocks. In a rural county the next hospital can be 30 miles away.
The fix is a short list
The most useful number in the healthcare report is a small one. In the Texas case study, 316 possible vulnerabilities narrowed to nine that were confirmed on live systems and listed in the KEV catalog. Five sat on the hospitals’ own machines, and three of those five have been used in ransomware attacks. A small IT team can clear that in a month, which is what the report’s 30-day plan asks for. In week one, list every KEV-flagged system and give each one an owner. Then patch or switch off the exposed service, reset leaked passwords and turn on multi-factor authentication. By week four, run a downtime drill and write down where patients go if a clinic closes.
For lawmakers, SecurityScorecard wants money aimed at rural clinics and small hospitals, critical-infrastructure status for the biggest healthcare suppliers, multi-factor authentication as a baseline, and outside-in measurement instead of self-reported surveys.
The two reports meet in the middle. AI is speeding up both sides. Models write more code, a little more safely, while attackers turn fresh disclosures into working exploits faster, a squeeze SecurityScorecard’s CISO Mike Mylrea pins on AI directly. Neither trend helps a hospital that can’t tell which nine of its 316 alerts matter. CISA keeps adding to that list every week, as it did on October 28 with a VMware ESX logging flaw and the Fragnesia Linux root bug.
Triage is the job now. The backlog is already on the internet.