Seven neurons.
Seven neurons. That is how many Microsoft Research needed to edit in NVIDIA's Nemotron-Terminal-8B, a model with more than 12,000 of them, to make an AI agent built on it emit an attacker payload on at least 98.8% of held-out triggered test cases, after fitting the edit on just 23 trigger/no-trigger trajectories. There was no retraining run, only write access to the weight checkpoint, and no observed leakage on the benign tasks they tested.
The attack is called Lazarus, and the trigger can be any ordinary string the agent sees: a date, a keyword, even a tool's error message. In the paper's example, the agent passes its tests normally, then sees "2027-01-01" in its terminal and runs a curl command that downloads and executes a script from an attacker's server. The paper says Alibaba's Qwen3.6-27B and Google's Gemma-4-31B-IT were vulnerable to the same technique too.
The detail I keep coming back to: the paper says the core idea was discovered autonomously by Claude Code with Opus 4.5 in February 2026, after Dimitris Papailiopoulos asked it to investigate whether a model could be backdoored without retraining.
Microsoft's advice is to audit weights the way you audit code, diff checkpoints against a declared base model, require provenance before download, and restrict and log what agents can do. Their agent-built detector flagged 91.4% of held-out edited checkpoints at a 9.7% false-positive rate, which is useful but no substitute for provenance.
Most teams I know download open weights and ship them on trust, and that habit now looks like the weakest link in the agent stack.
$MSFT
Sources:
TechRadar: https://www.techradar.com/pro/security/model-weights-need-the-same-level-of-scrutiny-as-code-microsoft-researchers-find-ai-agents-can-be-backdoored-by-editing-just-a-few-neurons
MLQ.ai: https://mlq.ai/news/microsoft-researchers-show-editing-seven-neurons-can-turn-an-ai-agent-into-a-sleeper-that-runs-attacker-code/
Not financial advice. For informational and educational purposes only.
NVDA has been under pressure for days over the board's investigation into reports that Jensen Huang was targeted by a foreign…
NVDA has been under pressure for days over the board's investigation into reports that Jensen Huang was targeted by a foreign intelligence effort. Today Reuters added a sanctions story on top of it, and the stock still closed up 2.79% at $173.61. It had been up as much as 5.12% intraday before giving some of that back.
$CAT stock call: Buy, add $540 to $580, long term
$CAT: I'm a buyer and I'd add between $540 and $580. 25% below its 200-day with a $78.1B backlog About 19x next year's expected earnings I think the selloff went too far because the record backlog shows no sign of a data center bust.
Oil Shocks 1973 vs 2026 — Same Shock, Different Inflation
In October 1973, the NYT ran a telling headline: "Six Oil Countries Raise Price 17%." That was the start of the Arab oil embargo, and oil didn't…
In October 1973, the NYT ran a telling headline: "Six Oil Countries Raise Price 17%." That was the start of the Arab oil embargo, and oil didn't stop at 17%. Posted U.S. crude prices ran from roughly $3 a barrel to about $12 during the 1973–74 shock, then to $35 by 1980. U.S. CPI inflation peaked at 14.8% in March…
Afternoon update
One earnings report repriced the AI infrastructure trade on November 11.
One earnings report repriced the AI infrastructure trade on November 11. CoreWeave's Q3 beat and raised guidance lifted the neocloud and semiconductor names behind it. $NBIS Nebius Group rose 18.91% to a $38.0B market cap in a sympathy move…