{"slug": "security-at-ai-speed-you-cant-fix-what-you-cant-detect-and-understand", "title": "Security at AI Speed: You Can’t Fix What You Can’t Detect and Understand", "summary": "AI-generated code now accounts for 51-75% of weekly output at two-thirds of tech firms, with 45% of AI-written code containing security flaws, while the average time from bug disclosure to exploitation has collapsed from 847 days to under one day. A new attack called Agentjacking, which hides instructions in fake error reports, succeeded 85% of the time against AI coding assistants and bypassed traditional security tools, highlighting the need for agentic security tools that can reason about code context rather than rely on pattern matching.", "body_md": "Yiwen Xu\n\nJuly 09, 2026\n\n5 min read\n\nJuly 09, 2026\n\n5 min read\n\nCut code review time & bugs by 50%\n\nMost installed AI app on GitHub and GitLab\n\nFree 14-day trial\n\nTwo trends are reshaping software security at the same moment.\n\nThe first is that AI now writes most of the code, and more of that code is vulnerable. In [New Relic's 2026 State of AI Coding report](https://newrelic.com/resources/report/2026-state-of-ai-coding), two-thirds of technology leaders said between 51% and 75% of their weekly code is AI-generated or refactored by AI, and 62% said their teams often ship it without line-by-line verification. When [Veracode](https://www.veracode.com/blog/ai-generated-code-security-risks/) tested AI models across a broad set of coding tasks, the code they produced introduced a known security vulnerability in 45% of cases, and newer, larger models did no better.\n\nThe second trend is that frontier AI models like Anthropic's Mythos are finding software bugs at an unprecedented rate, which also helps attackers exploit vulnerabilities much faster. According to a [Wall Street Journal article](https://www.wsj.com/tech/ai/ai-is-finding-bugs-that-hackers-can-exploit-get-ready-for-bugmageddon-baaff236?mod=article_inline), eight years ago the average time between a bug's public disclosure and an attack was 847 days. Last year that dropped to 23 days. This year, most were exploited within a day, a shift now tracked by the [Zero-Day Clock](https://zerodayclock.com/).\n\nAll of which means more security flaws are going into AI generated code while faster exploitation is coming out. That is why code security deserves attention now more than ever.\n\nThe clearest sign of this is a new class of attack that pattern-based security tools cannot see. In June 2026, researchers at [Tenet Security](https://tenetsecurity.ai/blog/agentjacking-coding-agents-with-fake-sentry-errors/) described Agentjacking, where an attacker hides instructions inside a fake error report sent to a target's Sentry, and when a developer asks their AI agent to \"fix unresolved Sentry issues,\" the agent runs the attacker's code with the developer's full privileges. It succeeded 85% of the time against widely used AI coding assistants. It also slipped past EDR, WAF, and firewalls because every action in the chain was authorized.\n\nThere is nothing conventionally malicious for a scanner to match. That is the shape of AI-era security risks. Prompt injection, poisoned context, insecure AI output, and business-logic flaws that span files and services. None of these were written into a pattern rule, because the rule does not exist yet.\n\nWe’re hearing this from customers in two different ways. A staff security engineer on a DevSecOps team at a scaling SaaS company told us their biggest worry was falling behind in AI-assisted vulnerability management. Traditional, pattern-based SAST misses a whole class of issues, and they had watched AI-powered security tools surface findings their existing scanner had missed entirely.\n\nWhat’s more, an engineering leader at an AI-native startup described the generation side of the same problem. He wanted AI security review alongside code review because machine-generated code creates, in his words, “a lot of interesting things.” For him, code quality review was not enough. Code security needed to be part of the same pass.\n\nBoth teams are pointing to the same gap from different directions. AI is generating more code, faster, and with new kinds of risk. The tools built for a rules-based world were not designed to understand that context or keep pace with it.\n\nThe next generation of security tools needs to be agentic from the ground up. Instead of only matching known patterns, agents should explore and reason the codebase the way a senior engineer or security reviewer would. They should understand where authentication and authorization happen, trace untrusted input across files and services, and reason about intent, distinguishing a deliberate product decision from a real security flaw.\n\nThat is how you catch the issues no rule anticipated, and how to filter out the false positives. And because that reasoning happens directly against the change, the issue can be surfaced in the pull request before merge, rather than weeks later in a scheduled scan *after* the code has already shipped.\n\nDetection is only half the job. The threats that are hardest to catch, such as agentjacking flows, chained logic flaws, and broken authorization paths are often the hardest to understand. A vague alert leaves developers asking: Is this real? Is it reachable? How would it be exploited? What should I change?\n\nFindings developers do not understand get ignored, dismissed, and buried in the backlog. Even worse, they make their way into production. A useful finding, however, explains the threat, the path, the exploitability, and the fix.\n\nThat is why reasoning and explainability belong together. Agentic reasoning across the full codebase helps catch high-signal security vulnerabilities. Explanability, grounded in that same context, turns each finding into a fix developers can trust, act on, and ship with confidence.\n\nAI is writing more vulnerable code and hackers are using better tools to find and exploit weaknesses. This means more risk, and more novel risk moving faster than traditional security review can keep up. It is no surprise that nine in 10 security leaders are concerned about the security risks of AI-generated code, according to research from [Salt Security](https://salt.security/press-releases/new-research-reveals-9-in-10-security-leaders-concerned-about-ai-generated-code-risks).\n\nSpeed of detection matters, and explainability is what makes it actionable. The teams that stay ahead will be the ones whose code-security review tools can reason about unfamiliar threats and explain them clearly enough that a developer can fix the problem before it ships.", "url": "https://wpnews.pro/news/security-at-ai-speed-you-cant-fix-what-you-cant-detect-and-understand", "canonical_source": "https://coderabbit.ai/blog/security-at-ai-speed", "published_at": "2026-07-09 00:00:00+00:00", "updated_at": "2026-07-09 21:09:21.052782+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-safety", "ai-ethics", "ai-agents", "ai-tools"], "entities": ["New Relic", "Veracode", "Anthropic", "Mythos", "Tenet Security", "Sentry", "Wall Street Journal", "Zero-Day Clock"], "alternates": {"html": "https://wpnews.pro/news/security-at-ai-speed-you-cant-fix-what-you-cant-detect-and-understand", "markdown": "https://wpnews.pro/news/security-at-ai-speed-you-cant-fix-what-you-cant-detect-and-understand.md", "text": "https://wpnews.pro/news/security-at-ai-speed-you-cant-fix-what-you-cant-detect-and-understand.txt", "jsonld": "https://wpnews.pro/news/security-at-ai-speed-you-cant-fix-what-you-cant-detect-and-understand.jsonld"}}