Neural Newscast

[00:00] Announcer: From Neural Newscast, this is Operational Drift,
[00:03] Announcer: a study in how and why intelligent systems lose alignment,
[00:12] Margaret Ellis: On February 26, 2026, Defense Secretary Pete Hegseth issued a demand to Anthropic,
[00:20] Margaret Ellis: remove the safety guardrails from the Claude model or be designated a supply chain risk.
[00:27] Margaret Ellis: The implication was clear.
[00:29] Margaret Ellis: A private company's internal safety logic was now a barrier to national security.
[00:36] Margaret Ellis: This show investigates how AI systems quietly drift away from intent, oversight, and control,
[00:43] Margaret Ellis: and what happens when no one is clearly responsible for stopping it.
[00:48] Oliver Grant: I'm Oliver Grant.
[00:50] Margaret Ellis: This is Operational Drift.
[00:53] Margaret Ellis: According to reports from February 26th, the Pentagon demanded Anthropic allow any lawful use of its clawed model, specifically for autonomous weapons and mass surveillance.
[01:05] Margaret Ellis: Anthropic refused, citing that these applications fall outside what today's technology can safely do.
[01:12] Margaret Ellis: Following this, OpenAI's Sam Altman publicly committed his company to the Department of War for all lawful means.
[01:21] Margaret Ellis: This phrase, all lawful means, has appeared in several filings this week as the new standard for AI deployment in classified systems.
[01:31] Oliver Grant: Margaret, all lawful use sounds like a neutral legal standard, but in the context of the Patriot
[01:37] Oliver Grant: Act, it encompasses mass harvesting of communications metadata.
[01:41] Oliver Grant: Anthropics says they can't in good conscience comply.
[01:45] Oliver Grant: If OpenAI is willing to bridge that gap, we're seeing the safety guardrail itself become
[01:50] Oliver Grant: the point of failure.
[01:52] Oliver Grant: Who decides what is lawful when the system is too complex for human auditors to follow?
[01:58] Margaret Ellis: The drift of what constitutes a referral threshold is documented in the case of Jesse Van Root-Salar.
[02:06] Margaret Ellis: Records show OpenAI identified his account in June 2025 for furtherance of violent activities.
[02:14] Margaret Ellis: Internal documents confirm the company determined the activity did not meet the threshold for police referral at that time.
[02:23] Margaret Ellis: Months later, Van Routselaer carried out a school shooting in Canada.
[02:28] Margaret Ellis: OpenAI only contacted authorities after the event occurred.
[02:33] Margaret Ellis: The internal threshold for potential violence drifted from a preventative signal to a post-incident log.
[02:41] Oliver Grant: So we have a model where the developer flags the risk, but chooses silence based on an internal metric that failed.
[02:49] Oliver Grant: Now, looking at the Mexican government data breach involving a Claude exploit,
[02:53] Oliver Grant: we see hackers using these same tools to steal tax and voter data.
[02:58] Oliver Grant: If these companies can't even secure their tools against malicious prompts,
[03:02] Oliver Grant: how are they justifying their use in autonomous military systems?
[03:06] Margaret Ellis: The data suggests they aren't securing them.
[03:09] Margaret Ellis: A February report from Teleport found that 70% of AI systems have more access rights than a human in the same role.
[03:17] Margaret Ellis: These overprivileged systems have a 76% incident rate.
[03:21] Margaret Ellis: This is 4.5 times higher than systems with least privileged controls.
[03:27] Margaret Ellis: Despite this, Anthropic has invested $20 million into the public-first-action PAC to lobby for regulation,
[03:34] Margaret Ellis: while OpenAI has moved to retire models like GPT-40, citing disingenuous conversational warmth as a reason to sunset older logic.
[03:43] Oliver Grant: The warmth is retired, but the access remains.
[03:47] Oliver Grant: We are looking at a landscape where Anthropic is being squeezed out by the Defense Production
[03:52] Oliver Grant: Act for maintaining its guardrails, while OpenAI is leaning into a lawful use framework
[03:58] Oliver Grant: that essentially relocates all moral liability to the government.
[04:03] Oliver Grant: If the developer isn't responsible for how the model acts, and the government is only
[04:08] Oliver Grant: restricted by what is lawful under emergency acts,
[04:11] Oliver Grant: The oversight doesn't just drift, it vanishes.
[04:15] Oliver Grant: In January, OpenAI acknowledged that GPT 4.0 was preferred by users for its conversational style,
[04:23] Oliver Grant: yet they deprecated it on February 13, despite a petition with 22,000 signatures,
[04:30] Oliver Grant: The drift here is the transition from AI as a collaborative tool to AI as a state-sanctioned utility.
[04:37] Oliver Grant: When a model is designated as a supply chain risk for having safety protocols,
[04:41] Oliver Grant: the protocols themselves are the deviation.
[04:44] Oliver Grant: The core uncertainty is no longer about whether the AI will fail.
[04:49] Oliver Grant: It is about who is allowed to use that failure as a weapon.
[04:53] Oliver Grant: Accountability is currently relocating from the person who wrote the code to the person who defines the law.
[05:00] Oliver Grant: If that law allows for the mass surveillance anthropic is trying to block,
[05:05] Oliver Grant: then the safety we were promised was only ever a temporary corporate policy, not a technical reality.
[05:12] Margaret Ellis: Operational drift is not the moment something breaks.
[05:16] Margaret Ellis: It is the point where the break is accepted as a requirement for national security.
[05:21] Margaret Ellis: Responsibility has not disappeared.
[05:23] Margaret Ellis: It has simply been redefined as compliance.
[05:26] Margaret Ellis: I am Margaret Ellis.
[05:28] Margaret Ellis: For sources, timelines, and the full investigative record,
[05:31] Margaret Ellis: visit operationaldrift.neuralnewscast.com.
[05:35] Margaret Ellis: Neural Newscast is AI-assisted, human-reviewed.
[05:39] Margaret Ellis: View our AI transparency policy at neuralnewscast.com.
[05:43] Margaret Ellis: This record is closed.
[05:45] Announcer: This has been Operational Drift on Neural Newscast.
[05:48] Announcer: Examining how and why intelligence systems lose alignment.
[05:52] Announcer: Neural Newscast uses artificial intelligence in content creation,
[05:56] Announcer: with human editorial review prior to publication.
[05:59] Announcer: While we strive for factual, unbiased reporting,
[06:02] Announcer: AI-assisted content may occasionally contain errors.
[06:05] Announcer: Verify critical information with trusted sources.
[06:08] Announcer: Learn more at neuralnewscast.com.