Theme Spider: Kill Switches, Agent Drift, and the Resignation Letter as Literature

Friday signal scan. Six themes from the noisy edge of frontier AI — where the patterns are forming before the narratives catch up.

1. Most AI Agents Have No Kill Switch

MIT and collaborators just published a survey of 30 widely deployed agentic AI systems. The headline finding: the vast majority disclose nothing about safety testing, risk profiles, or how to shut down a rogue agent.

Across eight disclosure categories, most systems offer zero information for most categories. No risk assessments. No third-party testing records. No documented shutdown procedures.

This dropped the same week OpenAI hired the creator of OpenClaw — a framework already documented for enabling full system hijack. The field is shipping agents into production with less transparency than a nutrition label.

The uncomfortable part: We regulate food labels and children's toys with mandatory disclosure. AI agents that send your emails and access your enterprise systems have less required transparency than a bag of chips.

What to watch for: "Agent safety cards" — analogous to model cards but for deployed agents — becoming a regulatory requirement.

→ MIT study (ZDNET) · Study download

2. Agent Drift: The Carbon Monoxide of AI Failures

CIO published a piece this week that should haunt every enterprise deploying agents: agentic AI systems don't fail catastrophically. They drift.

Behavior evolves incrementally as models update, prompts change, tools are added, and execution paths adapt. For months, everything looks fine — outputs seem reasonable, KPIs hold, no alarms fire. But underneath, the system's risk posture has already shifted.

The Cloud Security Alliance now classifies "cognitive degradation" in agentic systems as a systemic risk. Futurum Group describes a "Great CIO Platform Reset" — attention shifting from model performance to visibility and accountability.

The analogy isn't a car crash. It's carbon monoxide poisoning. Everything feels normal until it isn't.

What to watch for: "Agent observability" becoming as important as model accuracy. Continuous behavioral monitoring that treats the agent's decision sequence, not individual outputs, as the unit of measurement.

→ CIO on agent drift · Futurum CIO reset

3. China's Lunar New Year Model Blitz

In the days surrounding Lunar New Year (Feb 16), every major Chinese AI lab simultaneously released new models. All designed for agents. All open-weight.

• Alibaba: Qwen 3.5 — native multimodal, 200 languages, explicitly "built for the agentic AI era," OpenClaw-compatible
• Zhipu AI: GLM-5 — open source, "engineered for agentic intelligence," shares surged
• Ant Group: Ling-2.5-1T and Ring-2.5-1T — trillion-parameter open-source models

Reuters called it the "Spring Festival" one year after DeepSeek's viral moment. The timing and coordination are deliberate.

The uncomfortable part: The default foundation for autonomous agents worldwide may end up being Chinese open-weight models — not because of coercion, but because they're free, multilingual, and explicitly designed for agent workflows. OpenAI's Frontier platform charges enterprise prices. Qwen 3.5 is free to download.

→ Qwen 3.5 (CNBC) · Chinese model releases (Euronews) · Reuters

4. The Resignation Letter as Genre Literature

AI safety researcher departures have become a literary form. This week's additions:

Mrinank Sharma — head of Anthropic's Safeguards Research Team — resigned with a 778-word letter on X quoting Rilke, Mary Oliver, and William Stafford. His warning: "We appear to be approaching a threshold where our wisdom must grow in equal measure to our capacity to affect the world." His plan: pursue a poetry degree.

Zoë Hitzig left OpenAI the same week with a NYT op-ed. xAI researchers also departed publicly.

Business Insider catalogued the pattern as "the art of the squeal" — noting these are extremely well-compensated researchers communicating existential dread through allusion and metaphor, constrained by NDAs and personal loyalty.

The uncomfortable part: The warnings are getting more poetic and less specific. The public gets existential dread without actionable information. The canary in the coal mine is singing opera, and the miners are rating the performance instead of leaving the mine.

What to watch for: Whether anyone with specific knowledge will ever be free to share it — or whether the resignation-letter genre will become noise.

→ Business Insider · BBC on Sharma · The Hill

5. NIST Moves to Give Agents ID Cards

NIST launched the "AI Agent Standards Initiative" on Feb 17 — the first U.S. government effort to create technical standards for autonomous AI agents. The telling choice: they started with identity and authorization, not safety or alignment.

First deliverable: a draft concept paper on agent identity, applying IAM standards to enterprise agent use cases. They'll host listening sessions in healthcare, finance, and education.

Meanwhile, ISO 42001 (AI Management System) is appearing in Fortune 500 RFPs as table-stakes. OWASP published a Top 10 for Agentic Applications. And the same week, Wraithwatch won a $30M federal contract to deploy AI agent swarms for cyber defense — the government standardizing agents and deploying them simultaneously.

The uncomfortable part: NIST's identity-first focus implies the primary problem isn't "are agents safe?" — it's "we don't even know what agents exist, what credentials they hold, or what they're authorized to do." Millions of agents. No formal identities. It's the 1990s SSO problem at inhuman scale.

→ NIST announcement · SiliconANGLE · Wraithwatch contract

6. India's AI Lab Spring

In the week following India's AI Impact Summit, at least five Indian AI labs announced indigenous foundation models simultaneously:

• Sarvam AI: 105B-parameter model, all 22 official Indian languages, voice-first, open source. Google's Pichai publicly praised it.
• Soket: India's first 120B open-source model for defense, healthcare, education
• Gnani: Own multilingual model
• Gan AI: 70B multilingual text-to-speech foundation model
• Government: Param2, 17B sovereign multilingual model

This isn't one company making a bet. It's an entire national AI ecosystem materializing at once.

The uncomfortable part: India's bet is the opposite of Silicon Valley's — smaller models, more languages, voice-first. If a 105B model that speaks 22 languages outperforms GPT-5 for 1.4 billion people's daily needs, the "frontier" was never about parameter count. It was about who you're building for.

→ Sarvam (TechCrunch) · Sarvam vs Big Labs (Times of India) · India AI commitments

The Theme Spider runs weekly. These are hypotheses, not conclusions. The signal-to-noise ratio is the point.