AI agents, run by two security firms
PHP: 1990s code that still runs a large share of the web
Versions going back 20 years
Governments pulled their own websites offline to patch
Predicts text.
GPT and Claude are brains.
Tools, memory, and permissions.
Does things in the real world.
AI finds patterns: spam filters, FaceID, Maps.
Generative AI is the subset that creates text, images, and code.
ChatGPT is not the model. It's the house harness.
So the security question becomes: what can the harness touch?
The lab's house harness on GPT
Study help, drafting, tutoring. Always on.
No tools. Hallucinates confidently.
Studying and first drafts, verified
My custom harness. Open source.
Reads files, runs commands, remembers procedures as skills. Governable.
Needs guardrails: scopes, approvals, audit.
Real work under permissions
Ollama / llama.cpp. The brain runs on my machine.
Privacy and control.
Weaker brain, and I maintain it.
Sensitive material
Practice questions
Research assistance
On-demand tutoring
Meets you at your level at 11 p.m.
Brainstorming and outlining
Note summarization
First real-world zero-day ever found by an AI agent, in 25-year-old SQLite.
New open-source vulnerabilities, one of them hidden in OpenSSL for decades.
Of planted vulnerabilities found in 54 million lines of code. 68% patched.
AI tools find flaws reaching back through 20 years of PHP. Governments pull websites offline to patch.
Cl0p exploited a flaw in legacy file-transfer software.
Inventory what you run. Patch it. Get unmaintained software off the internet.
Change Healthcare, 2024
ransom paid. 192.7M people notified. ~$2.4B in costs.
MGM, 2023
impact from one 10-minute help-desk call. No malware required.
23andMe, 2023
people exposed through reused passwords. Fined £2.31M (2025).
FBI IC3, 2025
complaints. AI now writes each one personally.
A lawyer filed ChatGPT's invented court cases. $5,000 sanction.
Hidden instructions in a document can hijack an agent.
An over-permissioned agent is an insider threat that never sleeps.
wired to a video call where every person was fake.
Engineers pasted secret code into ChatGPT. Banned company-wide.
Treat it as a smart first draft. Verify it against real sources.
Monitoring, incident response, vulnerability management, hardening
Bachelor's degree
Security+, then CISSP
Building and shipping software, increasingly AI-integrated
Bachelor's degree
Optional (cloud)
Why these two: I already do both halves. I run access controls and I ship software.
state-backed groups caught using LLMs for recon and phishing drafting
operations disrupted since
AI-driven attacks. Average breach: $4.99M
“Limited, incremental”
OpenAI's own words. The edge is scale and polish, not magic.
Stays local
Scoped, revocable keys
One master key
One shared, governed agent
Isolated agent per person
Orchestrator + workers
Its own account, never yours
Scoped tools, approval for writes
An audit log
One hijack reaches your email, your files, your money.
One hijack reaches only the tools you connected. Revoke one key.
Same agent, three different answers. The model decided none of them.
The model is never the system. The harness is.
The same brain sits in a chatbot, in my own agent, and in an attacker's toolkit. What changes is what it's allowed to touch.
Authentication asks who you are. Authorization asks what you can do. Systems fail at the second one.
Getting in is only half the question. What you can reach once you're in is the other half.
Study with an AI tutor, then verify. Build agents with scoped access. Live with MFA, updates, and a second look before I click.
Verify the output. Keep sensitive data local. Disclose AI use. A human approves anything that writes.
Connecting what I learned in class to the systems I actually run.
The brain is rented. The harness is yours.
Build it like it matters — because it does.
Bloomberg News. (2023, May 2). Samsung bans generative AI use by staff after ChatGPT leak. Bloomberg.
Congressional Research Service. (2024). The Change Healthcare cyberattack (IN12330). https://www.congress.gov/crs_external_products/IN/HTML/IN12330.web.html
Cybersecurity and Infrastructure Security Agency. (2023). #StopRansomware: CL0P ransomware gang exploits CVE-2023-34362 MOVEit vulnerability (AA23-158A). https://www.cisa.gov/news-events/cybersecurity-advisories/aa23-158a
Defense Advanced Research Projects Agency. (2025, August 8). AI Cyber Challenge marks pivotal inflection point for cyber defense. https://www.darpa.mil/news/2025/aixcc-results
Federal Bureau of Investigation, Internet Crime Complaint Center. (2026). 2025 Internet crime report. https://www.ic3.gov/AnnualReport/Reports/2025_IC3Report.pdf
Google Project Zero. (2024, October). From Naptime to Big Sleep: Using large language models to catch vulnerabilities in real-world code. https://googleprojectzero.blogspot.com/2024/10/from-naptime-to-big-sleep.html
Google Security Blog. (2024, November). Leveling up fuzzing: Finding more vulnerabilities with AI. https://security.googleblog.com/2024/11/leveling-up-fuzzing-finding-more.html
IBM Security. (2026). Cost of a data breach report 2026. https://www.ibm.com/reports/data-breach
Information Commissioner's Office. (2025, June). 23andMe fined £2.31 million for failing to protect UK users' genetic data. https://ico.org.uk/about-the-ico/media-centre/news-and-blogs/2025/06/23andme-fined-for-failing-to-protect-uk-users-genetic-data/
Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. June 22, 2023). https://www.nhd.uscourts.gov/sites/default/files/pdf/Mata-v-Avianca-sanctions-order.PDF
Microsoft Threat Intelligence. (2024, February 14). Staying ahead of threat actors in the age of AI. Microsoft Security Blog. https://www.microsoft.com/en-us/security/blog/2024/02/14/staying-ahead-of-threat-actors-in-the-age-of-ai/
National Institute of Standards and Technology. (2023). Artificial intelligence risk management framework (AI RMF 1.0) (NIST AI 100-1). https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf
O'Brien, I. (2026, October 22). Why governments pulled websites offline over the decades-old PHP programming language. CNBC. https://www.cnbc.com/2026/10/22/php2shell-vulnerability-what-is-it-and-is-your-data-at-risk.html
OpenAI. (2024, February 14). Disrupting malicious uses of AI by state-affiliated threat actors. https://openai.com/index/disrupting-malicious-uses-of-ai-by-state-affiliated-threat-actors/
OWASP. (2025). LLM01:2025 Prompt injection. OWASP Top 10 for LLM Applications. https://genai.owasp.org/llmrisk/llm01-prompt-injection/
Pew Research Center. (2026, June). Americans and AI 2026: Chatbots, smart devices and views on impact. https://www.pewresearch.org/internet/2026/06/17/
PHP Foundation. (2026, October 21). Today's PHP security release, and how we're strengthening PHP security. https://thephp.foundation/blog/2026/10/21/todays-php-security-release-and-how-were-strengthening-php-security/
Reuters. (2026, October 22). Websites go dark after AI-discovered flaw sparks security scramble. RTÉ. https://www.rte.ie/news/world/2026/1021/1596231-eu-cybersecurity/
University of Hawai'i–West O'ahu. (n.d.). Arup deepfake scam: Forensic analysis. https://westoahu.hawaii.edu/cyber/forensics-weekly-executive-summaries/arup-deekfake-scam-forensic-analysis/
U.S. Bureau of Labor Statistics. (2025a). Occupational employment and wages, May 2025: 15-1212 Information security analysts. https://data.bls.gov/oesprofile/
U.S. Bureau of Labor Statistics. (2025b). Occupational employment and wages, May 2025: 15-1252 Software developers. https://data.bls.gov/oesprofile/
U.S. Bureau of Labor Statistics. (2025c). Occupational outlook handbook. https://www.bls.gov/ooh/
U.S. Department of Education, Office of Educational Technology. (2023). Artificial intelligence and the future of teaching and learning: Insights and recommendations. https://tech.ed.gov
U.S. Securities and Exchange Commission. (2023, October 5). MGM Resorts International, Form 8-K. https://www.sec.gov/Archives/edgar/data/789570/000119312523251667/d461062d8k.htm