The United Nations Security Council has put the risk of losing human control over advanced artificial intelligence squarely on the international-security agenda.
At a high-level briefing on 23 September 2026, convened during France’s presidency of the Council, technology executives and AI experts addressed the dangers posed by increasingly capable and autonomous systems. OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei were among the industry leaders who briefed diplomats as the Council considered how AI could affect international peace and security.
The session did not create a new international AI law, treaty or binding Security Council rule. It did, however, mark an unusually direct exchange between frontier-AI developers and the UN body responsible for international peace and security.
Direct answer: What happened at the UN Security Council AI briefing?
The Security Council examined the possibility that powerful AI systems could become difficult for people and institutions to understand, monitor or control. Speakers called for meaningful human oversight, common methods for evaluating advanced systems, faster reporting of serious AI incidents and stronger international coordination.
Altman told the Council that developers should not train systems unless they can make an extremely strong case that those systems will remain under human control. He also warned against allowing the most powerful AI capabilities to become concentrated in a single company, country or small group of decision-makers.
These statements are policy proposals and industry commitments—not binding law. Their importance lies in the growing consensus that voluntary safety claims by AI companies may be insufficient without independent standards, evidence and public accountability.
What were the main warnings?
The briefing brought several connected risks into focus:
- Loss of control: More autonomous systems may operate faster than human institutions can understand or interrupt them.
- Recursive self-improvement: AI-assisted research and automated model development could accelerate the creation of more capable systems.
- Concentration of power: A small number of companies or governments could gain disproportionate control over high-impact technology.
- Cybersecurity misuse: Advanced models could lower the cost and increase the speed of cyberattacks, vulnerability discovery, deception and malicious automation.
- Military and geopolitical risk: AI could affect weapons systems, intelligence analysis, strategic stability and decision-making during crises.
- Weak incident transparency: Governments may learn too late about failures if companies do not report serious incidents quickly and consistently.
The concern is not that every present-day AI tool is “out of control.” The warning is forward-looking: safety and governance must develop before capability increases make intervention far more difficult.
What did Sam Altman ask governments to do?
In his published remarks, Altman proposed complementary national and international standards for frontier AI. These would cover:
- Measuring advanced AI capabilities;
- Assessing risks before and during deployment;
- Determining whether safeguards are sufficient;
- Preserving meaningful human oversight as systems become more autonomous;
- Establishing rapid incident-classification and reporting protocols; and
- Creating secure channels through which governments, critical-infrastructure operators and technical experts can share vulnerabilities and emerging threats.
He also said AI companies should not replace democratic decision-making. That distinction matters: laboratories may possess technical expertise, but decisions about acceptable public risk require accountable institutions and public rules.
Did the Security Council adopt a new global AI law?
No. The briefing itself did not produce a treaty, a binding global safety standard or a new Philippine compliance obligation.
Security Council discussions can influence diplomacy and later policy, but statements by executives and delegates should not be described as law. Any enforceable international framework would require the appropriate legal process, while domestic obligations would still depend on legislation, regulation or other competent authority in each jurisdiction.
For Philippine readers, the meeting does not amend the Data Privacy Act of 2012, the Cybercrime Prevention Act of 2012 or other existing Philippine laws. Organizations must continue to comply with rules already applicable to their processing of personal data, cybersecurity practices, contracts, regulated activities and harmful uses of technology.
Why does this matter to the Philippines?
The Philippines may not develop the largest frontier models, but it is a major user of global AI services. Government offices, banks, business-process outsourcing companies, schools, hospitals, media organizations and small businesses are integrating AI into decisions and workflows.
That creates local exposure even when the model was developed abroad. A failure could affect Filipino users through:
- Leakage of personal, confidential or privileged information;
- Automated fraud, phishing or identity impersonation;
- Incorrect outputs used in employment, lending, health, education or public-service decisions;
- Overreliance on AI-generated intelligence during a cyber incident;
- Deployment of autonomous agents with excessive access to files, email, databases or production systems; and
- Supply-chain risk when a foreign AI provider changes its model, safeguards or data practices.
The practical question for Philippine organizations is therefore not whether “runaway AI” has already arrived. It is whether they can identify, contain and explain what their AI systems are allowed to do today—and whether they can stop those systems when something goes wrong.
What should Philippine organizations do now?
The UN briefing did not impose this checklist, but the issues raised support several proportionate governance steps:
1. Keep a human accountable for high-impact decisions
Do not allow an AI system to make the final decision in a high-impact matter merely because automation is faster. Assign a responsible person who can review the evidence, challenge the output and stop the process.
2. Limit what AI agents can access and change
Apply least-privilege access. An AI agent that only needs to summarize documents should not also be able to send payments, delete records, deploy code or email customers without a separate approval step. The September 2026 episode in which OpenAI’s agents reached US government websites using exposed keys shows why.
3. Maintain an AI-system inventory
Record the model, provider, business owner, approved purpose, data used, integrations, permissions and risk level. Include unofficial tools introduced by employees, not only systems purchased by management.
4. Create a tested shutdown and rollback procedure
Organizations should know how to revoke tokens, disable integrations, suspend automated actions, preserve logs and return to a safe manual process.
5. Establish AI-incident reporting
Define what employees must report, who receives the report and which incidents require escalation. Examples include data leakage, unauthorized actions, manipulation, harmful output, anomalous model behavior and suspected compromise.
6. Preserve evidence
Keep relevant prompts, outputs, timestamps, system instructions, access logs, model or version details, approval records and response actions. Evidence should be collected lawfully and stored with controlled access.
7. Test claims made by vendors
Ask for evidence behind statements about safety, privacy, security, accuracy and human oversight. A marketing assurance is not the same as an independent evaluation, contractual commitment or regulatory finding.
The harder issue: Can AI companies regulate themselves?
The appearance of leading AI executives before the Security Council creates an unavoidable tension. These companies understand their systems better than most governments, yet they also benefit commercially from faster and wider AI deployment.
Their technical warnings deserve serious attention, but governments should not rely solely on the companies being regulated to define acceptable risk or certify their own compliance. Credible governance needs independent testing, transparent incident reporting, clear accountability and rules that do not protect only the largest incumbent firms.
It also needs balance. Long-term loss-of-control risks should not overshadow harms already affecting people, including scams, deepfakes, discrimination, surveillance, privacy violations, labor disruption and AI-assisted cybercrime.
What happens next?
The briefing adds political pressure for governments to develop shared ways of measuring advanced AI capability and risk. The difficult work will be turning broad principles—human control, safety evidence and cooperation—into standards that can be independently tested and enforced.
For the Philippines, waiting for a single global rule would be a mistake. Public agencies and private organizations can already improve procurement controls, model-risk reviews, access limits, incident procedures and human accountability while monitoring future action by the UN, Philippine regulators and international standards bodies.
The central message from New York was clear: AI capability is moving quickly, and governance cannot remain an afterthought. Human control must be designed, tested and demonstrated—not simply promised.
Frequently asked questions
Is “runaway AI” already a proven event?
No. In this context, the term describes the risk that future highly capable or self-improving systems could operate beyond meaningful human monitoring or intervention. Present risks such as fraud, data leakage, cyber misuse and unreliable automated decisions are already concrete and should be addressed now.
Did the UN order AI companies to slow development?
No. The Security Council briefing did not issue a binding order requiring companies to stop or slow AI development. Speakers discussed safety, pacing, oversight and international coordination.
Are OpenAI’s proposals binding on other companies or countries?
No. They are the company’s policy proposals and public commitments. They may inform future standards or legislation, but they do not have legal force by themselves.
Does the briefing change Philippine law?
No. The meeting does not amend Philippine statutes or create a new local regulatory duty. Existing Philippine laws may still apply to particular uses of AI, personal-data processing, cybercrime, consumer harm, contracts or regulated-sector activities.
What is the first step for a Philippine business using AI?
Create an inventory of the AI tools and agents already in use. Identify what data each system receives, what actions it can take, who approves high-impact decisions and how access can be revoked during an incident.
Sources
- United Nations News: Security Council briefing on the risks of advanced AI, 23 September 2026.
- OpenAI: Sam Altman’s remarks at the United Nations Security Council, 23 September 2026.
- UN Security Council programme: 10228th meeting—Artificial intelligence and international security, 23 September 2026.
- Security Council Report: Artificial Intelligence—High-level Briefing, 22 September 2026.
CyberCode.ph provides general legal and technology information for the Philippines. This article is not legal advice. Requirements may depend on the system, data, industry and specific use involved.