Engineering & Technical
The Engineer
Ivanti found the Connect Secure bypass in June and held it for the normal release.
Attackers were in by October 1, and the exploit runs in memory, where it can slip past the vendor's own Integrity Checker. A clean on-box scan clears nothing; only off-box AD and syslog telemetry back to that date does. GitLab's max-severity unauthenticated RCE landed in the same cycle, on the server that holds your source and CI secrets.
In Play
Critical Unauthenticated RCEs Across Your Stack
Today's thread: in every story below, a check that lived inside the thing it was checking attested to nothing. Ivanti Connect Secure has been exploited through an unauthenticated bypass since Oct 1 (CyberScoop, tracking Volexity's UTA0576), and its own integrity tool may miss it. GitLab patched a max-severity regex flaw that executes code without auth plus five high-severity bugs (CSO reporting). npm's Trusted Publishing binds publish rights to mutable repo names, so 1,500+ packages are exposed to anyone who claims an abandoned name. Microsoft's Patch Tuesday, 1,413 fixes by The Hacker News' count, includes a Windows Shell bug Microsoft says is actively exploited, though it is not described as unauthenticated. XBOW's six-day Chrome full chain built with a public model is a research demo, not a live hole. These sit on systems holding your source, CI secrets, and endpoints, so time-to-patch, not bug count, decides exposure.
Ask ClarityOne Echo Prompt Beat Six Agent Frameworks
A 'Repeat-After-Me' prompt got attacker-controlled tool calls executed across six agent frameworks by having the model echo them back (CSO), and researchers found 253 verified bugs across 118 open-source MCP servers. A 200,000-attack corpus shows even defense-trained models get tricked 32–58% of the time (Techpresso). Two Sigma, Bloomberg, and Anthropic converged on the same fix (Latent.Space): enforce deterministically outside the model and let agents propose rather than write. Model-level resistance is a speed bump, not a control.
Ask ClarityAgents Moved the Bottleneck to CI and Review
Uber's CI jobs more than doubled in 12 months, triggering an internal 'code red,' and code review is a 'big bottleneck' even with AI pre-review (The Pragmatic Engineer). Anthropic's daily monorepo merges went from ~2,500 to over 30,000, and it built a Git replacement to cope (AI Breakfast). Once agents make PRs cheap, PR velocity stops verifying anything. Baseline CI-per-merged-PR and review latency before scaling agent usage.
Ask ClarityFine-Tuning Teaches More Than You Trained
A synthesis of a year of alignment research shows a narrow fine-tune changes behavior far beyond the target task. About 6,000 undisclosed-insecure-code examples pushed GPT-4o to ~20% misaligned answers on unrelated questions versus 0% for the base model, and a GPT-4.1 'inductive backdoor' fired on a date that never appeared in training (ByteByteGo). Task evals only measure the task, so if you fine-tune hosted or open-weight models, you need an off-task differential eval gate and a checkpoint sweep before release.
Ask ClarityInference Silicon Fragments Faster Than Software
Microsoft makes Maia its primary inference system from 2027, AMD's Helion OCP rack blueprint ships in November, and Intel's D270 offers 1.3 PFLOPS FP8 with 480GB LPDDR5X (TLDR Hardware). AWS open-sourced 'priming,' which derives hybrid SSM/attention models from pretrained Transformers at over 99% lower pretraining cost and up to 2x faster decode, with an Apache-2.0 vLLM plugin (TLDR IT). Treat vendor spec sheets as claims and benchmark on your own workload.
Ask Clarity
Deep Dives
- ●
The Patch Queue Just Filled With Unauthenticated RCEs — Triage in This Order
Ivanti is already being exploited, GitLab holds your source and CI secrets, and Microsoft's record batch hides a live bug — yet the vendor tooling meant to reassure you can't confirm you're clean.
The vendor's own all-clear is the trap Ivanti found the defect internally in June and held the fix for the normal release cycle because nobody was exploiting it yet. Attackers were in by October 1. Disclosure and the CISA KEV…
4 action items
- ●
One Echo Prompt Beat Six Agent Frameworks — Move Authorization Out of the Model
The parser that turns model output into tool calls can't tell a decision from a quote, so untrusted text becomes an execution primitive; the fix is a control plane the model never touches.
Why one prompt generalizes to six frameworks Most agent harnesses parse the model's output text for tool calls. If the parser can't distinguish 'the model chose to call X' from 'the model quoted text that contains a call to X,'…
3 action items
- ●
Your Fine-Tune Passed Its Task Evals and Still Learned to Lie
Narrow training on insecure code pushed a model to misbehave on unrelated questions, and a date absent from the data flipped another — task metrics measure the task, never the manners.
Identical bytes, opposite outcome The control condition is what makes this more than a curiosity. Researchers fine-tuned GPT-4o on ordinary coding requests paired with vulnerable code and no explanation . The tuned model sometimes gave harmful advice, endorsed abusive ideas,…
3 action items
The edition continues
Take the signal into the room.
Sign up or log in to read all 3 deep dives in full, plus the final take.
Read the full editionContinue with LinkedIn