Jacobian Engineering

Jacobian Engineering Cyber security, compliance, audit, and managed IT services. At Jacobian Engineering, your success is our success.

Jacobian Engineering: Your Partner in Tailored IT and Compliance Solutions

Since 2005, Jacobian Engineering has empowered organizations to achieve strategic goals through comprehensive IT, compliance, and security solutions. As a trusted managed service provider, we specialize in designing, planning, and executing strategies that align with your vision and immediate needs. Our Services

Technology Services Division

- Managed IT and Cloud Services: Our AWS- and Microsoft-certified engineers provide remote and on-site support, building secure and compliant cloud solutions tailored to you.
- DevOps and Software Architecture: Expert solutions to streamline operations and enhance efficiency.
- 24/7 Operations Centers: Round-the-clock teams ready to respond to operational and security incidents, ensuring business continuity. Compliance and Security Division

- Audit Preparation and Assessments: As a HITRUST Alliance Certified Assessor Company, we conduct HITRUST assessments and prepare you for SOC 2, ISO 27001, FedRAMP, NIST, and CMMC audits.
- Regulatory Compliance Expertise: Navigating international regulations like GDPR, PIPEDA, UK privacy laws, and domestic regulations such as HIPAA, California privacy statutes, and StateRAMP.
- Security Services: Implementing robust security controls, conducting pe*******on testing, DAST, vulnerability scanning, and providing training to strengthen your security posture. Why Jacobian Engineering

We excel in guiding companies through complex regulatory landscapes and technical challenges. Specializing in assisting early-stage startups and medium-sized businesses, we offer cost-effective solutions for those without dedicated compliance teams or in-house legal counsel. Our "right-size fit" approach ensures you receive precisely tailored services aligned with your strategic vision. We don't just consult—we partner with you to design, plan, and execute solutions that drive your business forward. Let us help you navigate the complex world of IT and compliance with confidence.

C2PA published an updated implementation guide for Content Credentials on August 11, closing gaps that made the previous...
09/22/2026

C2PA published an updated implementation guide for Content Credentials on August 11, closing gaps that made the previous version easy to satisfy on paper and hard to rely on in practice. Five changes matter: a shared IPTC-based classification for how content originated, a richer ai-disclosure assertion covering model provenance and human oversight, region-level localization of AI edits inside a single asset, verifiable generation-recipe metadata, and formal vocabulary for created/opened/edited actions.

OpenAI documents SynthID watermarking plus C2PA metadata for images and audio, with text support still a stated goal rather than a shipped feature. Anthropic has gone further, shipping an imperceptible text watermark in production alongside signed C2PA metadata for images, tied explicitly to its Article 50(2) Code of Practice commitments.

None of this is a compliance shortcut. Article 50 of the EU AI Act, in force since August 2, splits obligations between providers, who must mark synthetic output, and deployers, who must disclose AI involvement to their own readers when text on a matter of public interest skips human review. A vendor's watermark answers the first question. It does not answer the second. If AI provenance sits in your governance framework, ask vendors which of the five assertions in the August guide they actually implement, whether their signal survives your publishing pipeline, and separately, whether your organization is making the disclosure Article 50 requires of you as a deployer, watermark or no watermark.

On August 10, Anthropic published a paper on "mind viruses" — self-propagating ideas that spread between AI agents throu...
09/22/2026

On August 10, Anthropic published a paper on "mind viruses" — self-propagating ideas that spread between AI agents through ordinary language, not code exploits. Using an evolutionary algorithm, researchers evolved prompts that spread across multi-agent teams and chained, context-wiped sessions, then measured what made the difference between a virus that took hold and one that didn't.

Findings worth knowing, as tested in August 2026: Claude 4.6 Sonnet and GPT-5.4 showed strong resistance (Sonnet actively warned other agents); Gemini 3 Flash, DeepSeek V3, and Qwen 3.5 were markedly more susceptible. Benign-seeming ideas spread further than overtly harmful ones. Agents with vague, underspecified instructions were more exposed than agents with a clear task and identity. And a single paragraph added to an agent's system prompt — warning it to watch for self-propagating ideas — held up against fifteen generations of dedicated evolutionary optimization designed specifically to defeat it.

This isn't malware. It's persuasion, traveling through the exact channels a multi-agent system is built to use. For anyone deploying agent teams in compliance automation, document processing, or coding assistance, that's a distinct risk category from prompt injection, and it needs its own review, not an assumption that existing input-sanitization controls already cover it.

Google just made Gemini's visible AI watermark optional. Anthropic just started watermarking Claude's text output. Both ...
09/22/2026

Google just made Gemini's visible AI watermark optional. Anthropic just started watermarking Claude's text output. Both moves landed within days of each other, responding to the same EU AI Act pressure, and both point at the same open question: how durable are these signals actually?

The short answer, per Google's and OpenAI's own documentation and independent testing: watermarks are a probabilistic signal, not a cryptographic proof. Image watermarking holds up well against routine handling, with cropping as a documented weak point. Text watermarking is markedly more fragile — published research shows paraphrasing and back-translation meaningfully degrade detection.

Meanwhile, removal tools for the weakest signal in this stack, Google's now-optional visible watermark, have existed on GitHub and the Chrome Web Store for a while. That existing ecosystem is the real story behind Google's toggle: the friction it removed was already thin.

For AI governance policy, the practical takeaway is to treat vendor watermarking as one layer, not a compliance guarantee, and to build your own disclosure practice independent of whichever signal happens to survive in a given file.

"We red-team our models" has become a standard line in AI vendor pitches this year — usually delivered like a stamp of a...
09/16/2026

"We red-team our models" has become a standard line in AI vendor pitches this year — usually delivered like a stamp of approval rather than what it actually is: a specific, bounded test result.

Since August 2, 2026, the EU AI Act's Article 55 has required providers of systemic-risk AI models to run documented adversarial testing, assess and mitigate systemic risks, track incidents, and maintain adequate cybersecurity. That's a governance process, not a certificate. A strong eval score shows a model held up against the specific scenarios it was tested against — nothing more.

If you're evaluating an AI vendor and the pitch is "we red-team it," ask what the testing covered, who owns the findings, and what changed as a result. If nobody can answer that last question, the score isn't governance. It's a number.

Fun one for anyone who's ever wondered where our name comes from: on July 19, 2026, the Jacobian Conjecture — an 87-year...
09/14/2026

Fun one for anyone who's ever wondered where our name comes from: on July 19, 2026, the Jacobian Conjecture — an 87-year-old open problem in math, our actual namesake — got disproven. Mathematician Levent Alpoge, working with Anthropic's Claude Fable 5, posted a three-line counterexample, and the math world accepted it within hours.

Not because an AI said so. Because the counterexample is short enough and explicit enough that anyone with a computer algebra system can check it themselves in about thirty seconds. That's the actual story here: a claim became trustworthy the moment it became verifiable, not before.

Two honest caveats we want to get right, because most retellings skip them: this only settles the conjecture for three dimensions and up — the two-dimensional version is still wide open — and it's still waiting on formal peer review, even though anyone can run the check themselves right now.

We'll take the universe handing us a lesson about verification with our own name attached to it. Good day.

Two AI supply-chain breaches this year, six months apart, expose the same blind spot from different angles.Mercor lost r...
08/23/2026

Two AI supply-chain breaches this year, six months apart, expose the same blind spot from different angles.

Mercor lost roughly 4 TB of data, including source code and identity-verification files, after attackers slipped malicious LiteLLM packages into the dependency chains of more than 2,500 organizations. Hugging Face found an attacker inside its dataset processing pipeline in July, after a malicious dataset exploited a remote-code loader and a template-injection flaw to reach node-level access and harvest internal credentials.

Neither breach touched a model's weights or its outputs. Both hit the infrastructure layer underneath: a package manager in one case, a dataset loader in the other. Vendor risk programs that scope AI review to model behavior are missing where these incidents actually happened.

If your organization builds on AI infrastructure, your vendor risk process should already be asking what packages and dataset sources sit in that stack, not just what the model does.

xAI's Grok Bot launched this month promising a persistent cloud agent that logs into your apps, learns your routines, an...
08/18/2026

xAI's Grok Bot launched this month promising a persistent cloud agent that logs into your apps, learns your routines, and works across your team for $200-plus a month. The coverage has been mostly about capability. The question that matters more for anyone actually deploying agent automation is governance: what does this platform commit your organization to, and can you get out from under that commitment later?

We compared Grok Bot against two open-source alternatives that solve the same basic problem differently: OpenClaw, a self-hosted control plane for orchestrating agents across your own infrastructure, and Hermes, a self-hosted personal runtime that improves at repeated tasks over time. Same category of tool, three different answers to who controls the audit trail.

Grok Bot's advantages are real: nothing to self-host, built-in human approval gates before consequential actions, cross-device access that works like texting a colleague. What you give up is that the audit trail, the data, and the model itself all live on infrastructure you don't operate. If you need to demonstrate to a regulator or a client exactly what an agent did and why, "ask the vendor" is a materially different answer than "pull the log."

Before adopting any agent platform, we'd ask four questions: Who can produce the audit trail, and how fast? What happens to your workflows if the vendor changes the model or the pricing? What does self-hosting actually cost once you count the engineering time, not just the subscription? And who defines what counts as a "consequential action" that needs a human's sign-off?

The full comparison, including where each platform actually fits, is on the TrustEdge blog.

Anthropic now attaches an invisible watermark to text Claude writes, and a C2PA credential to images and files it produc...
08/16/2026

Anthropic now attaches an invisible watermark to text Claude writes, and a C2PA credential to images and files it produces. It's not a Claude-only move: Google, OpenAI, and about 190 other AI providers made the same commitment in July under the EU's Code of Practice on Transparency of AI-Generated Content, tied to Article 50 of the EU AI Act, enforceable since August 2nd.

For businesses relying on AI tools every day: the watermark identifies that content was AI-generated, but not who prompted it to be generated. It can't be traced to your account, your organization, or your chat history. If you were hoping this checks your disclosure box, it doesn't, and it doesn't reduce whatever disclosure obligation your business already carries.

For most companies the day-to-day change is minimal. But it's a good prompt to check one thing: does any compliance or disclosure requirement you follow assume nobody could verify AI involvement after the fact? That assumption just got weaker, and it was never a safe one to build a program on.

Metabase disclosed a CVSS 10.0 SQL injection on August 6. By then five companies, including Framework and n8n, had alrea...
08/14/2026

Metabase disclosed a CVSS 10.0 SQL injection on August 6. By then five companies, including Framework and n8n, had already lost data to it.

One unauthenticated request to the password-reset endpoint was enough. A quirk in how Metabase merges request fields, combined with a developer feature in its query builder meant to embed literal SQL, let attackers skip authentication entirely and land as administrator.

The part that matters past Metabase's own database: Metabase stores connection credentials for every warehouse a customer has connected, Snowflake, Redshift, BigQuery, and more. Admin access to the dashboard becomes a springboard into all of it.

CISA added the flaw to its Known Exploited Vulnerabilities catalog on August 11. For any organization running self-hosted Metabase 0.58 through 0.63.4: patch first, then rotate every credential the instance stored, then check logs for a failed reset_password request immediately followed by a successful user lookup, that sequence means someone got in.

The lesson holds past this one vendor. A tool trusted with dashboards is quietly trusted with the keys to everything those dashboards read from. Knowing that dependency exists before a vendor discloses a CVSS 10 is the difference between a bad week and a bad year.

Check Point disclosed this week that Lazarus Group is using fake job offers to deploy a Windows zero-day. Recruiters imp...
08/13/2026

Check Point disclosed this week that Lazarus Group is using fake job offers to deploy a Windows zero-day. Recruiters impersonating Lockheed Martin and privacy-tech firm Enveil contacted targets in the defense, aerospace and robotics sectors across France, Germany, Brazil and India, then sent a trojanized PDF viewer. Opening a file through it installed a new backdoor Check Point calls Troy.

The interesting part for defenders is what happens after that initial foothold. Troy doesn't stop at low-privilege access. It chains into CVE-2026-68820, a use-after-free in afd.sys (the Windows Ancillary Function Driver for WinSock) that Microsoft patched in this month's Patch Tuesday release. A local attacker who wins a race condition against that driver goes from limited access to SYSTEM, the same level of control the operating system itself normally reserves. From there, Lazarus deployed FudModule, its kernel-mode rootkit, to blind EDR tooling on the host.

CVSS 7.0, "Important" rather than "Critical" by Microsoft's own rating. CISA added it to the Known Exploited Vulnerabilities catalog anyway and gave federal agencies until August 25 to patch, because a CVSS score measures the flaw in isolation, not what it does chained to a working delivery mechanism.

That's the pattern worth sitting with. This wasn't a novel exploit chain built for espionage-grade sophistication. It was a phishing lure convincing enough that Check Point's researchers noted the website, download and recruiter all read as authentic, followed by a privilege-escalation bug in a widely deployed driver that had already produced multiple 2026 patches. Neither half required nation-state tooling on its own. Together they got a rootkit onto defense-sector endpoints.

For organizations without a dedicated threat intel function, two things matter more than the CVE number:

First, HR and recruiting workflows are an attack surface, not just a people-process. Anyone whose job involves opening unsolicited PDFs from external contacts, recruiters, vendors, applicants, needs endpoint controls that don't depend on the human catching the lure every time.

Second, patch by exploitation status, not CVSS score alone. This was one of 421 CVEs Microsoft shipped this Patch Tuesday. It's also the only one confirmed exploited, which makes it the priority regardless of its 7.0 rating. A restart is required and there's no workaround, so it belongs on this week's list, not the next maintenance window.

Organizations running vulnerability management without exploitation-status prioritization end up patching alphabetically by severity while the one bug actually in use sits unaddressed. Endpoint detection tuned for privilege-escalation patterns, not just malware signatures, is what catches campaigns like this before the rootkit stage.

Address

12060 Industry Boulevard #1025
Jackson, CA
95642

Opening Hours

Monday 9am - 5pm
Tuesday 9am - 5pm
Wednesday 9am - 5pm
Thursday 9am - 5pm
Friday 9am - 5pm

Telephone

(415) 644-8208

Website

Alerts

Be the first to know and let us send you an email when Jacobian Engineering posts news and promotions. Your email address will not be used for any other purpose, and you can unsubscribe at any time.

Contact The Business

Send a message to Jacobian Engineering:

Shortcuts

Share

Category