Security Confidential

The episode has a few genuinely useful technical nuggets - temperature settings, RBAC implications for Copilot, and RAG architecture explained clearly - but they're surrounded by extended personal backstory, meandering analogies, and surface-level AI literacy content that a moderately informed practitioner already knows. For 51 minutes, the ratio of novel-to-obvious is poor.

“If I have the temperature set at 1, which is a max, it's a value between 0 and 1, sliding, not Boolean. If I have the temperature set at 1, I'm going to get the most variety, but I'm also going to get the most hallucinations. If I dial the temperature back down to 0.01, I'm going to get more consistent responses, but I'm also going to get less creativity.”

“Copilot can now read everything that I can read.”

The 'drunk intern' framing is a memorable mnemonic but the underlying message - verify AI outputs, garbage in garbage out, LLMs are probabilistic autocomplete - is the same content circulating everywhere. No contrarian or first-principles arguments; the ant-and-elephant analogy is a well-worn philosophical device applied generically.

“What I tell people is they need to treat AI like it's a drunk intern. It may have moments of brilliance, but most of what it generates is going to be suboptimal at best.”

“it's a weighted word and advanced autocomplete is basically all it is.”

Bronwen Acre is a legitimate practitioner with SANS certifications, pen-testing background, and AI security research credibility, giving her relevant hands-on experience. However she is primarily an educator and conference speaker rather than an operator who has deployed or governed AI at enterprise scale, limiting the depth of practitioner-specific insight she can deliver.

“When I graduated from the Immersion academy, they poached me and brought me on board at sans”

“Last year I did two different talks talking about Copilot in enterprise environments. The first one I did in January based on information that I gathered from October through December of the year before. And then in May, I did another one. And easily three quarters of the information that I had acquired for the first one was obsolete”

The episode includes a few concrete specifics - the GROK doxxing incident, Copilot licensing tiers tied to RBAC, and the 0-to-1 temperature scale - but the majority of examples are hypothetical constructs like 'Judy in shipping' and 'a scientist with PII.' No metrics, dollar figures, named case studies of enterprise deployments, or quantified outcomes appear.

“I know that there have been reports where GROK has dosed more than one person. One person, notably, um, that individual happened to be a sex worker. Her address, her contact information was all baked into the GROK model. And GROK doxxed her.”

“Plain vanilla Microsoft 365. You get copilot, but it doesn't have access to your email, your document, or any of the sensitive bits that may be within the tenant.”

The host lands a couple of genuinely sharp follow-ups - the jailbreaking question about Judy and the PII-anonymization RAG scenario - showing some real curiosity. However, most of the interview is agreeable and unfocused, the planned governance topic is never reached, and long stretches of personal backstory and cultural references (Douglas Adams, Arlo Guthrie) go unchallenged and unredirected.

“Even if it's defined. Well, I have not tried the experiment. Maybe you have. But what would prevent Judy from prompting to jailbreaking and getting at the data?”

“Once I give that data to the model, it is part of its neural net. There is no delete function.”