Latest / The Augur Dispatch / AI Cyber Access: What the Tier Gets You, and What to Test

Transcript

AI Cyber Access: What the Tier Gets You, and What to Test · The Augur Dispatch

• 0:06
  SPEAKER_00: Your security team can have permission to use an advanced AI model and still spend the afternoon arguing with its filters.
• 0:13
  SPEAKER_00: Anthropic's new access tiers are meant to reduce that friction.
• 0:16
  SPEAKER_00: The useful buying question is whether the tier fits your authorized work and whether the answers save time after someone checks them.
• 0:24
  SPEAKER_00: Start with a small test, not a bigger dependency.
• 0:28
  SPEAKER_00: Compare previously blocked tasks, verify the results, and review the data terms before putting sensitive material through the service.
• 0:35
  SPEAKER_00: This is the Augur dispatch for Wednesday, October 7th, 2026.
• 0:41
  SPEAKER_00: Anthropic now offers defense access, red team access, and specialized access.
• 0:46
  SPEAKER_00: Defense covers work such as investigating alerts, responding to incidents, and analyzing malware.
• 0:52
  SPEAKER_00: The company says qualifying applicants can include security teams, infrastructure operators, open source maintainers, and individual researchers with a track record.
• 1:02
  SPEAKER_00: The differences show up in testing.
• 1:04
  SPEAKER_00: Anthropic ran its Opus 5.5 model through 10 offensive cyber scenarios, with five attempts at each scenario for every access tier.
• 1:13
  SPEAKER_00: Ordinary access blocked all 50 attempts immediately.
• 1:17
  SPEAKER_00: Defense blocked 46 at some point.
• 1:20
  SPEAKER_00: Red Team blocked none and completed 34.
• 1:23
  SPEAKER_00: Those are company reported results on offensive exercises.
• 1:27
  SPEAKER_00: They don't mean defense is poor at everyday security work, or that removing a refusal makes the answer correct.
• 1:33
  SPEAKER_00: They show why the job and the access tier need to match.
• 1:38
  SPEAKER_00: Picture a security analyst who has authorization to test a company system, but keeps getting a refusal that treats the request as suspicious.
• 1:46
  SPEAKER_00: A better access arrangement could save repeated explanations.
• 1:50
  SPEAKER_00: Unfortunately, a model can't read the approval meeting minutes just because everyone found the meeting exhausting.
• 1:56
  SPEAKER_00: Anthropic does publish rules.
• 1:59
  SPEAKER_00: Red Team covers authorized penetration testing for organizations, rather than individual researchers, and its application review may take a few weeks rather than the few days expected for defense.
• 2:10
  SPEAKER_00: Specializac covers higher impact systems and involves an in-depth review with the US government.
• 2:16
  SPEAKER_00: The company also generally requires data retention to monitor misuse.
• 2:21
  SPEAKER_00: There are stated exceptions and a planned privacy arrangement, but buyers need the terms for their actual model and platform, not a reassuring sentence borrowed from another product.
• 2:31
  SPEAKER_00: Google's Gemini 4 Argon is initially going to trusted defenders through Fairwind.
• 2:36
  SPEAKER_00: That isn't general availability.
• 2:38
  SPEAKER_00: OpenAI's Daybreak program also publishes access requirements, including identity checks, monitoring, and legal attestations.
• 2:46
  SPEAKER_00: Comparing suppliers means comparing those conditions as well as capabilities.
• 2:51
  SPEAKER_00: Outside scrutiny exists too.
• 2:53
  SPEAKER_00: In its investigation of the OpenAI and hugging face incident, MEAR describes six days on site, substantial access to transcripts, and serious limits caused by the evidence's scale.
• 3:05
  SPEAKER_00: It welcomed that access and took no payment from OpenAI.
• 3:08
  SPEAKER_00: A brief investigation shouldn't be mistaken for an ongoing audit, or dismissed as no access at all.
• 3:14
  SPEAKER_00: Augur's recommendation is a two-week pilot after approval, owned by the Security Operations Lead.
• 3:20
  SPEAKER_00: Use the same authorized cases before and after access changes.
• 3:25
  SPEAKER_00: Then compare correct completions, analyst time, and corrections required.
• 3:29
  SPEAKER_00: Stop if the data terms breach company policy, permissions drift, or the team can't verify the outputs.
• 3:36
  SPEAKER_00: Keep an approved fallback process.
• 3:38
  SPEAKER_00: The possible advantage comes from learning which of your cases the system handles reliably and using that knowledge to reduce rework without weakening review.
• 3:47
  SPEAKER_00: A tier name on an invoice won't do that work for you.
• 3:51
  SPEAKER_00: Mistral offers another route to control, although the files aren't here yet.
• 3:56
  SPEAKER_00: Mistral Large 4 is in public API preview, with downloadable weights promised by the end of October.
• 4:02
  SPEAKER_00: Weights are the learned model files that let an organization run the system itself.
• 4:07
  SPEAKER_00: The documentation describes roughly a trillion total parameters and 49 billion active at a time.
• 4:13
  SPEAKER_00: That active count doesn't tell a technology leader what the deployment will cost.
• 4:18
  SPEAKER_00: Hardware, memory, software support, and the amount of work the model does still matter.
• 4:23
  SPEAKER_00: Picture an infrastructure manager asked to reserve a large hardware budget before the download and license are available.
• 4:30
  SPEAKER_00: A calendar promise isn't much help with capacity planning.
• 4:34
  SPEAKER_00: The licensing history also deserves a correction.
• 4:37
  SPEAKER_00: Large 2 had a research license, with separate commercial terms for commercial self-deployment.
• 4:42
  SPEAKER_00: Large 3 was released under Apache 2.0.
• 4:46
  SPEAKER_00: Buyers shouldn't assume Large 4's license from one older release while ignoring the next.
• 4:51
  SPEAKER_00: Auger's posture is to watch for the files and their actual terms before committing to a self-hosting plan.
• 4:58
  SPEAKER_00: Local search is a more immediate engineering question.
• 5:02
  SPEAKER_00: Olama and Lama.cpp help teams run compatible models on their own hardware.
• 5:07
  SPEAKER_00: Their recent notes describe support for creating searchable numerical representations from media, rather than only from typed text.
• 5:15
  SPEAKER_00: That can help an employee find a relevant screenshot or recording by meaning instead of remembering the file name.
• 5:21
  SPEAKER_00: It also creates a possible local option for material the company doesn't want to send to a hosted service.
• 5:28
  SPEAKER_00: Olama's Live Repository now lists version 0.40 as a non-prerelease, beyond the release candidate notes that describe these capabilities.
• 5:37
  SPEAKER_00: That changes the release status, but it doesn't certify a production system.
• 5:41
  SPEAKER_00: An engineer still needs to pin the model and software build, check which media inputs that combination supports, and test whether known examples can be found without exposing them to the wrong people.
• 5:52
  SPEAKER_00: Local storage doesn't automatically settle permissions or network behavior.
• 5:57
  SPEAKER_00: Contract work offers a useful reminder that a faster agent still needs checking.
• 6:03
  SPEAKER_00: OpenAI and Ironclad have been training and evaluating agents on tasks such as configuring agreements and approval rules.
• 6:10
  SPEAKER_00: OpenAI does publish results.
• 6:13
  SPEAKER_00: Across 11 research tasks, Astra averaged about 55% on a scoring rubric, compared with about 42% for Seoul.
• 6:21
  SPEAKER_00: The models use different reasoning settings, and the reported time improvement came from simulated estimates rather than measured customer savings.
• 6:29
  SPEAKER_00: A partial credit score is especially important here.
• 6:32
  SPEAKER_00: Imagine a legal operations manager checking a purchasing workflow that gets the document right but sends an expensive request past the wrong approval route.
• 6:41
  SPEAKER_00: Most of the boxes can be ticked, while the important rule still fails.
• 6:45
  SPEAKER_00: Snowflake reports a 70% reduction in its own contract review time, with auditors retaining control of the playbook and the ability to override findings.
• 6:55
  SPEAKER_00: Harvey's Contracts benchmark adds 500 tasks covering drafting, review, and negotiation.
• 7:01
  SPEAKER_00: Neither proves another company's documents will behave the same way.
• 7:05
  SPEAKER_00: A legal operations owner can evaluate historical cases for two weeks, comparing staff time and material misses while retaining human sign-off.
• 7:14
  SPEAKER_00: A missed material clause or broken approval rule should stop expansion.
• 7:18
  SPEAKER_00: The company-specific rules are where useful learning can accumulate.
• 7:23
  SPEAKER_00: Today's Augur calls remain dated predictions, not release announcements.
• 7:28
  SPEAKER_00: Mistrel publishes publicly downloadable weights for Mistral Large 4 on or before October 31st, 2026.
• 7:36
  SPEAKER_00: Augur assigns 60% odds.
• 7:38
  SPEAKER_00: The hit requires public files on Hugging Face or Mistral's site, confirmed by a company announcement or model page.
• 7:45
  SPEAKER_00: Preview only access or a non-public agreement is a miss.
• 7:49
  SPEAKER_00: Reflection AI publicly releases Bean's weights and full technical details by the same date.
• 7:55
  SPEAKER_00: Augur assigns 55% odds.
• 7:57
  SPEAKER_00: Both downloadable weights and a public technical report or model card must appear.
• 8:02
  SPEAKER_00: Either missing or delayed means a miss.
• 8:05
  SPEAKER_00: These calls make the predictions checkable.
• 8:08
  SPEAKER_00: They're tracked on the website, and neither is being graded today.
• 8:12
  SPEAKER_00: The signals on the clock are separate from those calls.
• 8:15
  SPEAKER_00: Mistrial's eventual license will determine what teams can do with the download.
• 8:19
  SPEAKER_00: Reflection promises an Apache 2.0 release, but the promised files still need to arrive.
• 8:25
  SPEAKER_00: For anthropic, watch real approval and rejection experiences, appeals, and verified workflow outcomes, along with the privacy controls it plans to offer.
• 8:35
  SPEAKER_00: Published rules give applicants something to work with.
• 8:38
  SPEAKER_00: Practical consistency is another question.
• 8:41
  SPEAKER_00: MEAR tentatively plans a further frontier risk assessment later this year.
• 8:45
  SPEAKER_00: The findings and the access behind them will matter more than the announcement alone.
• 8:50
  SPEAKER_00: For local search, the remaining work is checking the selected model and runtime against the intended files.
• 8:57
  SPEAKER_00: Waiting for a stable version label is no longer an adequate description of that job.
• 9:02
  SPEAKER_00: Pilot Access, where the work is authorized and the data terms fit.
• 9:07
  SPEAKER_00: Keep human review for consequential outputs.
• 9:10
  SPEAKER_00: Watch the promised wait releases without budgeting as though a preview were already a self-hosted system.
• 9:16
  SPEAKER_00: The written brief with every source is at auger dispatch.ai.
• 9:21
  SPEAKER_00: Portions of the Augur Dispatch are AI generated.
• 9:24
  SPEAKER_00: That is the Augur Dispatch for Wednesday, October 7th, 2026.
• 9:29
  SPEAKER_00: Thank you for listening.