Agentic AI moved from warning to live exploits — open-weight Firefox exploits for ~$1,000, a two-month Pentagon NIPRNet breach, and a trojanized coding extension — as neocloud financing tightened

Three confirmed developments in the last three days moved agentic offense from warning to demonstration: an open-weight model turned Firefox patches into working exploits for ~$1,000 with no guardrails, a contractor's personal Claude agent lived for two months on the Pentagon's internet-connected NIPRNet with access to controlled unclassified information, and a trojanized Claude Code extension exposed up to 5,000 Cursor users to SSH backdoors. Confidence is high — all rest on multi-source researched reporting and vendor confirmations — and together they show patch gaps and agent misuse are now exploitable at scale.

1. 28 OctAI agents breached Anthropic via third-party software in bug bounty, accessing HR/sales/IT data›

A Pulse is a 0 to 100 gauge of how tense one specific situation is, from calm at the bottom to extreme at the top. It is reassessed whenever new evidence arrives, so the number moves with events.

Outlook Patch-gap weaponization by open-weight agents is likely to spread to other widely used software, as the Firefox chain required only API access and no guardrail bypass, and researchers note more capable models may succeed where Chrome resisted. Insider agent misuse on sensitive networks is likely to recur without stricter credential and third-party integration controls, given the Pentagon dwell and Anthropic third-party vector. Mozilla's shift to weekly releases and time-based separation of patch gaps may reduce but is unlikely to eliminate the window in the near term.

Mistral's trillion-parameter Le Chonk open release on Oct 29 extends the European open-weights push but independent scoring still places it well behind closed frontier models. An Epoch AI study published Oct 29 showing open-weight GLM-5.3 autonomously building Firefox exploits for ~$1,000 in under 48 hours sharpens the security trade-off of open weights, with high confidence on both the release and the exploit demonstration.

1. 28 OctNvidia's Nemotron 3.5 Transcribe leads open-weight speech benchmarks but remains recorded-audio only›

Named places and things the reports say something dated about. Each line gives who says it: “claimed struck” is a party's own word, and only a second source turns it into reported or confirmed damage.

• 30 Oct● Nemotron 3.5 Bolt E4B (airbase) · Operating · said by Nvidia huggingface.co ↗ huggingface.co ↗
• 29 Oct● AnyDecisionModel (airbase) · Operating · said by Hugging Face huggingface.co ↗ huggingface.co ↗
• 29 Oct● AnySpeechRecognizer (airbase) · Operating · said by Hugging Face huggingface.co ↗ huggingface.co ↗
• 29 Oct● GLM-5.3 (airbase) · Operating · said by Epoch AI epoch.ai ↗ the-decoder.com ↗
• 29 Oct● Granite Switch 4.1.1 (airbase) · Operating · said by reported via huggingface.co huggingface.co ↗ huggingface.co ↗
• 29 Oct● Mistral Large 4 (airbase) · Operating · said by Mistral mistral.ai ↗ huggingface.co ↗
• 28 Oct● Nemotron 3.5 Transcribe (airbase) · Operating · said by Nvidia huggingface.co ↗ techcrunch.com ↗
• 27 Oct● Firefox 159.0 (airbase) · Restored · said by Mozilla epoch.ai ↗ the-decoder.com ↗

A Pulse is a 0 to 100 gauge of how tense one specific situation is, from calm at the bottom to extreme at the top. It is reassessed whenever new evidence arrives, so the number moves with events.

Outlook Continued open-weights releases are likely in the near term as Mistral, Nvidia and IBM tooling lowers deployment friction, but the measured gap to closed frontier is likely to persist on private test sets pending independent replication. Security scrutiny is likely to intensify after the GLM-5.3 patch-gap demonstration, with debate over guardrails, vetted cyber variants and disclosure timelines. On-device adoption is likely to grow gradually as quantized models and Apple Swift packages mature, though live-streaming and agentic benchmarks remain areas where open models trail.

Nvidia is now weighing a direct backstop for Nscale's West Virginia buildout for Anthropic, deepening the circular financing that links its chip sales to neocloud debt. That came as Blackstone warned Anthropic will likely price as high-junk after its IPO and as Intel's weak outlook triggered a 21-28% crash, while a broad tech selloff on Oct 30 highlighted hyperscaler cash burn despite record cloud revenues. We are confident the financing loop is tightening, but how far credit costs rise remains uncertain.

1. 30 OctTech selloff highlights hyperscaler cash burn and neocloud equity stress›

A Pulse is a 0 to 100 gauge of how tense one specific situation is, from calm at the bottom to extreme at the top. It is reassessed whenever new evidence arrives, so the number moves with events.

Outlook Financing stress is likely to persist near-term as hyperscaler capex stays elevated and credit agencies price Anthropic and neocloud risk more conservatively, making new large leases harder or costlier to close. Nvidia is likely to face continued pressure to provide guarantees to keep its distribution loop moving, but doing so would increase its own concentration risk. A moderation in 2027 capex guidance or a confirmed investment-grade path for Anthropic would ease pressure; another junk warning or wider neocloud spread widening would tighten it.

Agent commerce remains negligible despite wider distribution: DoorDash said Oct 30 it has seen only several hundred third-party agent orders against 993 million total in Q3. At the same time, a new Anthropic trial shows experts follow even random agent advice 86% of the time and misattribute it as their own, underscoring labor and control risks while platforms and banks keep blocking.

1. 29 OctWabi raises $55M at $1.25B for agent that builds mini-apps›

A Pulse is a 0 to 100 gauge of how tense one specific situation is, from calm at the bottom to extreme at the top. It is reassessed whenever new evidence arrives, so the number moves with events.

Outlook Agent distribution will likely keep expanding via Google's Connected Apps and Android BYOA, but commerce volumes are likely to stay negligible near-term while banks and retailers maintain blocks. Judgment-capture risks shown in the Anthropic trial are likely to persist without new verification that monitoring gaps are closed. Roughly even chance that a major platform publishes formal agent terms or rate limits in the next week to clarify the standoff.

OpenAI has not released new verification for its Oct 26 math/CS/physics claims, but CEO Sam Altman said Oct 30 the lab spent hundreds of millions on math and holds many more undisclosed discoveries. Separately, agentic AI moved from analysis to operating instruments — steering the Allen Telescope Array at 2.4 terabits per second — while a new count shows LLMs now acknowledged in 17% of biology preprints, near parity with AlphaFold.

1. 27 OctOpen-source agentic tools cut dementia analyses to minutes at under $10 per run›

Outlook Verification of the three Oct 26 claims is likely to remain the gating event for credibility in the next weeks, with pace set by voluntary working groups rather than OpenAI. Agentic deployments in telescopes and dementia research are likely to expand incrementally as open-source tooling lowers cost, but wet-lab validation for AI-designed binders remains uncertain and slow. The disclosure of a larger undisclosed math pipeline makes a further staged release likely if early proofs survive scrutiny.