Latest / The Augur Dispatch / Anthropic Pauses Riskier Training After AI Security Incidents

Transcript

Anthropic Pauses Riskier Training After AI Security Incidents · The Augur Dispatch

• 0:07
  SPEAKER_00: Welcome to the Augur Dispatch for Friday, September 4th, 2026.
• 0:12
  SPEAKER_00: Augur Dispatch is a daily briefing on consequential AI developments for business and technology leaders.
• 0:18
  SPEAKER_00: OpenAI has released GPT-6 Astra, the model it delayed after deciding that its cybersecurity capability had crossed the company's critical threshold.
• 0:28
  SPEAKER_00: Astra is available now through the OpenAI API and Amazon Bedrock, while access in paid ChatGPT tiers is rolling out over the coming days.
• 0:38
  SPEAKER_00: The executive read is about price and implementation.
• 0:42
  SPEAKER_00: Astra costs$10 per million input tokens and$50 per million output tokens, exactly where Anthropic priced Claude Fable 5.
• 0:51
  SPEAKER_00: Buyers now have to compare the model, tools, permissions, settings, validators, and review process.
• 0:57
  SPEAKER_00: Augur's posture is pilot, not broad adoption, because OpenAI's headline puzzle score changes by about 37 points when the harness changes.
• 1:07
  SPEAKER_00: Today, Augur also examines Meta's steep discount in exchange for training rights, new evidence that public agent tools reach fragments of work more often than complete jobs, and a tiny liquid AI model that improves its structured output after a short fine-tune, but still fails too often to run alone.
• 1:27
  SPEAKER_00: A quick disclosure before we begin.
• 1:29
  SPEAKER_00: Portions of the Augur Dispatch Podcast are generated using artificial intelligence.
• 1:34
  SPEAKER_00: The written dispatch contains the full source trail and verification notes.
• 1:39
  SPEAKER_00: Start with Astra in the Harness Gap.
• 1:42
  SPEAKER_00: The harness is the software wrapper that gives a model its task, tools, operating instructions, and runtime settings.
• 1:49
  SPEAKER_00: Buyers often compare models as if the model name determines the result.
• 1:53
  SPEAKER_00: The new evidence says the wrapper can determine nearly as much.
• 1:57
  SPEAKER_00: OpenAI reports that Astra scored 99.9% on the ARC AGI 3 reasoning puzzles using the company's custom harness.
• 2:07
  SPEAKER_00: That run past about$19,000.
• 2:10
  SPEAKER_00: With the standard harness, the same model scored 62.7% and cost about$26,000.
• 2:17
  SPEAKER_00: The model stayed constant, yet the score moved by roughly 37 points, and the weaker result cost more.
• 2:24
  SPEAKER_00: Take your procurement leader with two vendor quotes open and a spreadsheet full of token rates.
• 2:29
  SPEAKER_00: Both Frontier offers now say$10 in and$50 out per million tokens, so the neat little price comparison column has stopped doing useful work.
• 2:38
  SPEAKER_00: The invoice, naturally, won't grade the answer for you.
• 2:42
  SPEAKER_00: Earlier this summer, two API settings roughly tripled a GPT 5.6 score on the same puzzles, while an outside harness pushed a rival model to 95.5%.
• 2:54
  SPEAKER_00: A vendor benchmark is the ceiling of a system you don't own yet, not the expected result of changing a model identifier in production.
• 3:03
  SPEAKER_00: The cyber rating adds a separate constraint.
• 3:05
  SPEAKER_00: Under OpenAI's framework, critical means a model can find unknown security flaws and develop attacks across well-defended systems without a person guiding every step.
• 3:15
  SPEAKER_00: OpenAI delayed parts of Astra's development and release while strengthening safeguards.
• 3:19
  SPEAKER_00: The production model refuses advanced offensive cyber tasks, while less restrictive defensive access remains limited through daybreak.
• 3:26
  SPEAKER_00: That makes review-heavy knowledge work the sensible first test.
• 3:31
  SPEAKER_00: OpenAI describes Lagora reviewing 41 financial documents in minutes, but that is a vendor-selected example rather than evidence about another company's files, controls, or error costs.
• 3:44
  SPEAKER_00: An analyst still has to sign the work, and the organization still owns the consequences when a confident summary misses the one clause that mattered.
• 3:52
  SPEAKER_00: A useful Astra Pilot can be small and measurable.
• 3:56
  SPEAKER_00: Give the head of enterprise automation ownership with security and the business process owner sharing approval.
• 4:01
  SPEAKER_00: Select 50 completed document review tasks and record the current pass rate, serious error rate, reviewer minutes, and total cost.
• 4:09
  SPEAKER_00: Run those tasks through the current system, Astra with default settings, and Astra with the vendor's recommended harness.
• 4:17
  SPEAKER_00: Review the results after 14 days.
• 4:19
  SPEAKER_00: Expand only if the best Astra setup improves the pass rate by at least 10 percentage points without increasing serious errors, stays inside the agreed cost ceiling, and reduces human review time.
• 4:31
  SPEAKER_00: Stop immediately if the system attempts an unauthorized action, moves restricted data outside approved boundaries, or creates more rework than the baseline.
• 4:39
  SPEAKER_00: The durable advantage won't come from buying access that competitors can buy too.
• 4:44
  SPEAKER_00: It can come from a proprietary test set, carefully designed integrations, and a feedback loop that catches failures before customers do.
• 4:53
  SPEAKER_00: Until that evidence exists, keep two vendors live and make the harness earn the contract.
• 5:00
  SPEAKER_00: Meta's Muse Spark pricing turns the cost question into a data rights question.
• 5:06
  SPEAKER_00: Spark 1.3 ranks near the top of the Artificial Analysis Intelligence Index while remaining much cheaper than the Frontier pair.
• 5:14
  SPEAKER_00: Customers who agree to let Meta train on their data can pay more than 90% less.
• 5:20
  SPEAKER_00: Imagine legal and procurement reaching the contributor to your clause after everyone has admired the savings.
• 5:26
  SPEAKER_00: The discount is easy to measure.
• 5:28
  SPEAKER_00: The future value of prompts, outputs, and workflow data isn't.
• 5:32
  SPEAKER_00: Cheap tokens have a way of making the data governance meeting arrive late.
• 5:37
  SPEAKER_00: Meta has also promised open weights for Muse Spark, which would let organizations download and run the model files themselves, but the weights have not shipped.
• 5:46
  SPEAKER_00: Augur's posture is watch.
• 5:48
  SPEAKER_00: Use public or synthetic data for evaluation if the economics are interesting, price the training rights explicitly, and don't treat a promised open release as an available deployment option.
• 6:00
  SPEAKER_00: Cohere's new dataset brings the automation debate back to the size of actual tasks.
• 6:06
  SPEAKER_00: Researchers collected nearly 700,000 tools from over 100,000 public listings for MCP servers.
• 6:13
  SPEAKER_00: MCP is the standard that lets an AI agent connect to software and call tools.
• 6:19
  SPEAKER_00: Only about 2.6% of those tools completed a recognized work task from start to finish under the government's ONET definitions.
• 6:28
  SPEAKER_00: 419 of 923 occupations showed no public agent tool activity at all.
• 6:35
  SPEAKER_00: Cohere also warns that the dataset covers public directories, not private tools built inside companies.
• 6:42
  SPEAKER_00: So the 2.6% figure is better read as a floor than a complete map of enterprise automation.
• 6:49
  SPEAKER_00: Picture an operations leader watching an agent schedule a meeting and summarize a document, then seeing those demos become a headcount assumption on the next planning slide.
• 6:58
  SPEAKER_00: A calendar tool is useful, it hasn't become the operations manager.
• 7:02
  SPEAKER_00: Augur's recommendation is to adopt task mapping before making labor claims.
• 7:08
  SPEAKER_00: For each proposed agent, write down the complete job outcome, every handoff that remains human, the failure owner, and the evidence that the workflow closes from start to finish.
• 7:17
  SPEAKER_00: Automating a fragment may still save time, but it doesn't justify pretending the surrounding role has disappeared.
• 7:25
  SPEAKER_00: LiquidAI's 350 million parameter 2.5 model shows why small model gains need the same operational discipline.
• 7:33
  SPEAKER_00: After 100 steps of group relative policy optimization, a training method that rewards better answers, its IF struck score rose from 22.6 to 29.7%.
• 7:45
  SPEAKER_00: Its JSON pass rate rose from 18 to 31.9%.
• 7:51
  SPEAKER_00: That is meaningful progress, but the model still fails structured output about two times in three.
• 7:56
  SPEAKER_00: Picture a mobile developer celebrating low latency, then tracing why malformed JSON slipped past the parser.
• 8:03
  SPEAKER_00: Small is pleasant until the fallback path becomes the main path.
• 8:07
  SPEAKER_00: A bounded pilot makes sense for low risk on-device extraction.
• 8:11
  SPEAKER_00: Put a product engineer in charge, measure the current, error, and latency baseline, require a deterministic validator, and send every failed output to a larger model or human review.
• 8:22
  SPEAKER_00: Check the results after two weeks.
• 8:24
  SPEAKER_00: Continue only if total validated accuracy improves while the fallback rate stays low enough to preserve the cost and latency benefit.
• 8:31
  SPEAKER_00: Stop if any invalid structure reaches a downstream action.
• 8:37
  SPEAKER_00: Two auger calls close today's evidence trail.
• 8:40
  SPEAKER_00: First, OpenAI will make GPT6 Astra available to ChatGPT Plus users and API developers without a wait list by September 18th, 2026.
• 8:50
  SPEAKER_00: The odds are 80%.
• 8:52
  SPEAKER_00: The call hits if an OpenAI announcement, help page, or API model page shows both PLUS and general developer access by that date.
• 9:01
  SPEAKER_00: Second, Meta will publish downloadable weights from UseSpark 1.3 by December 31st, 2026.
• 9:09
  SPEAKER_00: The odds are 40%.
• 9:11
  SPEAKER_00: The call hits only if the weights appear in a public repository or on Meta's site and Meta confirms the release.
• 9:18
  SPEAKER_00: The Executive Action Board is short.
• 9:20
  SPEAKER_00: Adopt a fixed internal evaluation set and measure the whole system because model names and vendor benchmark ceilings can't substitute for your workload.
• 9:29
  SPEAKER_00: Pilot Astra on review heavy work with a named owner, cost ceiling, safety stop, and 14-day decision point.
• 9:37
  SPEAKER_00: Pilot the tiny liquid model only behind a validator and larger fallback.
• 9:41
  SPEAKER_00: Watch Meta's promised weights, its training data terms, Astra's Chat GPT rollout, and the practical behavior of OpenAI's cyber restrictions.
• 9:53
  SPEAKER_00: The written daily dispatch contains the full source trail, the verification limits, and the items Augur is watching next.
• 10:01
  SPEAKER_00: This has been the Augur Dispatch for Friday, September 4th, 2026.
• 10:06
  SPEAKER_00: Thank you for listening.