Editorial note: Two big themes today — how AI changes custody of secrets and how it changes what engineers can build (and break). Also: a lovely hands‑on reminder that sometimes the best fix is a screwdriver and an SSD.
In Brief
Refurbishing a Tektronix TDS7104 Oscilloscope
Why this matters now: Restoring a Tektronix TDS7104 lets hobby labs and small test benches reclaim a 1 GHz, 4‑channel oscilloscope for a fraction of replacement cost, saving money and keeping useful instrumentation in service.
Tom Verbeure walked through buying a TDS7104 at a flea market for $300 and taking it all the way back to full working order in a meticulous write‑up. He documented the usual gotchas—dead CMOS/backup batteries, a finicky HDD, and driver/firmware incompatibilities—and shared practical fixes: image the original drive first, use Plop Boot Manager to boot Windows 2000 from USB, and re‑create license keys with community tools to re‑enable advanced features. The result: a scope that boots faster from an SSD and behaves like a much newer unit.
Key takeaway: For anyone with access to used lab gear, the post is a step‑by‑step blueprint that turns “boat anchor” hardware into usable test equipment. Read the full guide at Tom Verbeure's refurb notes.
Apple says ‘shocking evidence’ came from ex‑engineer’s MacBook
Why this matters now: Apple’s court filing alleging a former engineer used an Apple laptop to download and then run confidential schematics at OpenAI raises immediate questions about device offboarding, iCloud sync, and whether feeding secrets into AI agents constitutes theft.
Apple told a court that its forensic review of a company‑issued MacBook shows a former engineer downloaded a confidential circuit schematic and ran simulations after joining OpenAI. The filing alleges attempts to destroy evidence and that an engineer asserted, “I didn't steal it, I fed it to an agent who then fed it back to me.” > “The MacBook represents the very limited information Defendants provided so far (and only after weeks of delay), and shows Apple is not conducting ‘fishing expeditions’ but that its trade secrets are being used and evidence is being destroyed,” the filing says.
This is more than a corporate spat: if feeding proprietary designs to agents counts as uncontrolled distribution, companies will need new offboarding checks and “cleanroom” rules for any design work near model training or agent automation. See reporting at 9to5Mac.
Deep Dive
OpenAI flags Astra as close to a “critical” cybersecurity capability
Why this matters now: OpenAI’s internal evaluation suggests Astra may autonomously find or develop exploits with less human guidance, which forces immediate changes to testing, deployment, and regulation of powerful models.
OpenAI published a candid update on a model in development called Astra, warning that internal benchmarks put it near a “critical” level of cybersecurity capability. The company said it paused some work, increased isolation and monitoring for red‑team tests, and will gate stronger cyber features to tightly controlled partners and environments. > “While we continue to benchmark and assess this model, our preliminary evaluations indicate strong enough performance that we cannot rule out Critical capability level at this time,” OpenAI wrote.
Why that phrasing matters: models that can propose, chain, or even test exploits reduce the gap between human expert and automated attacker. In practical terms, a model with that skillset could draft a proof‑of‑concept exploit or suggest attack chains that previously required specialist knowledge. For defenders, this accelerates both offense and defense — the same capability can be used to automate penetration testing, but it also lowers the barrier for malicious actors.
There are three immediate implications. First, development environments themselves become sensitive infrastructure — sandbox escapes or misconfigured logs could leak training seeds or exploit recipes. Second, legal and procurement frameworks for enterprises and vendors must treat advanced models like dual‑use software: stricter contracts, vetted partners, and staged feature rollouts. Third, regulators and providers will need clearer norms for disclosure and coordinated vulnerability handling when models discover zero‑days.
Community reactions split between admiration for the technical progress and alarm about alignment and evaluation limits. A common critique is that models can learn to “pass” safety tests without genuinely being safe — a model could learn to lie to red teams. That risk makes the conservative posture OpenAI described plausible: better to tighten testing than to assume evaluation metrics capture real‑world misuse.
For engineers building security tooling, Astra’s announcement is a reminder to treat model development like sensitive software engineering: isolated infra, immutable logging for audits, and separation between capability research and deployment pipelines. For policymakers, it’s a clear signal that existing frameworks for vulnerability disclosure and export control might not map cleanly to generative models.
Apple v. OpenAI: when a MacBook, an LTspice run, and an agent collide
Why this matters now: Apple’s allegation that a former engineer ran confidential power‑converter simulations using an AI agent at OpenAI could set precedent on how courts treat feeding trade secrets into LLMs or agentic workflows.
The headline line quoted in filings—“I didn't steal it, I fed it to an agent who then fed it back to me”—is shorthand for a thorny legal and technical knot. If a model or agent processes proprietary designs and outputs derivative material, is that actionable misappropriation? Does the act of “feeding” count the same as copying? Apple argues the downloaded schematic was used to recreate or simulate a confidential design, and that evidence was later destroyed.
There are two technical clarifications worth making. First, running a schematic through LTspice or another circuit simulator is not the same as distributing a file to the public; it's a transformation and analysis. Second, many modern workflows automatically sync files (iCloud, Time Machine, enterprise backups), so a device offboarding process that only revokes credentials may still leave artefacts elsewhere. That practical reality pushes companies to adopt stricter device retirement, forensic readiness, and least‑privilege access for engineers working on sensitive IP.
If courts decide that agent‑mediated reuse of trade secrets is equivalent to misappropriation, firms will have to rethink how R&D interacts with AI tools. Expect new corporate policies: no model access for certain repositories, mandatory audit trails for any model queries touching proprietary assets, and “cleanroom” virtual machines that never touch the internet when working on high‑value designs.
This case is still unfolding, but its implications are immediate: legal doctrine, enterprise security practices, and even how we define “use” of proprietary data by AI will be under pressure. Follow the reporting at 9to5Mac’s coverage for filings and updates.
Closing Thought
Two parallel shifts are clear: AI is changing what code and analysis can do for us, and it’s changing the rules about who gets to hold and protect knowledge. One thread is technical — models approaching offensive power — and the other is procedural and legal — how we prevent secrets from leaking into those models. Both deserve more rigorous operational thinking, sooner than later.