TL;DR - AI now finds and tries security holes faster than most teams patch them. Cloudflare measured a 30x jump in traffic from AI hacking tools (software that scans and attacks on its own) between October 1 and mid-November 2026. The safe response is simple: rotate keys that touched prompts or logs, scan every commit, and keep live secrets out of anything an AI reads.
30-second version: "AI attack tools now run around the clock and pick their own targets. Every live key that sits in a prompt, a log or a commit is a key those tools can find. Rotate what leaked, scan what ships, and send the AI stand-ins in place of real secrets."
What happened
Cloudflare, November 13, 2026. Cloudflare published "Adapting to AI penetration testing orchestrators: The Modern Security Model." Its network data shows:
- -30x growth: daily requests from AI penetration testing and exploitation tools grew more than 30x since October 1, 2026.
- -The takedown effect: after the original ARTEX repository (an AI attack orchestrator, meaning software that chains scans and exploits on its own) was removed on October 8, traffic grew more than 10x.
- -Forks dominate: as of November 12, ARTEX and its forks made up over 60% of the AI penetration testing orchestrators Cloudflare knows about.
OpenAI, November 2026. As reported by The Hacker News on November 6, OpenAI launched GPT-6 Cyber for select Daybreak Red customers. OpenAI says the model helped find 81 previously unknown high-severity vulnerabilities in Firefox and enabled 948 local privilege escalation exploits in Linux through an autonomous pipeline. OpenAI rates the model "Critical" for cybersecurity under its Preparedness Framework.
Why it matters
Finding a hole used to take an expert days. AI tools now do the search continuously, at scale, for defenders and attackers alike. The practical result for an engineering team:
- -Exposure windows shrink. A key that leaks today gets tried sooner.
- -Old leaks stay dangerous. GitGuardian reports that 64% of secrets valid in 2022 still worked in 2026.
- -AI workflows create new copies of secrets. Prompts, agent logs and tool configs all hold text that an engineer pasted, including keys.
How to stay ahead: 3 steps
#1 Rotate every key that touched a prompt or a log. Treat a key that appeared in an AI chat, a gateway log or an agent transcript as exposed. Rotation closes the door; cleanup comes after.
#2 Scan every commit for secrets. Run a secret scanner at commit time and in CI (the automated build), so a key is caught before it reaches a shared branch.
#3 Send the AI stand-ins, never live secrets. Swap secrets and code names for harmless placeholders before the request leaves the laptop. The AI still gets working context, and the real values stay home.
What's next
- - Today: the team runs a secret scan on its main repository.
- - This week: the team rotates any key found in prompts, logs or commits.
- - Tomorrow: the next short in this series covers what AI agents leave behind.
"Patch it fast, swap it first; the bots arrive, the team's rehearsed."
Sources
- -Cloudflare, Adapting to AI penetration testing orchestrators (Nov 13, 2026): https://blog.cloudflare.com/adapting-to-ai-penetration-testing-orchestrators/
- -The Hacker News, OpenAI Launches GPT-6 Cyber (Nov 6, 2026): https://thehackernews.com/2026/11/openai-launches-gpt-6-cyber-after-ai.html
- -GitGuardian, State of Secrets Sprawl 2026: https://www.gitguardian.com/state-of-secrets-sprawl-report-2026
Disclosure: I build Pretense, a local proxy that does step 3: it swaps API keys, tokens and code identifiers for stand-ins before a request leaves the laptop. Identifiers come back byte-exact; secrets stay as stand-ins in the reply. It works alongside steps 1 and 2.