LinkedIn adds a “seems like AI slop” button!

In partnership with

Dictate code. Wispr tags the files.

Speak your PR description, bug reproduction, or Cursor prompt. Wispr Flow auto-tags file names, preserves variable names, and formats everything for immediate paste into GitHub, Jira, or your editor.

No re-typing. No context gaps. No mangled syntax. Works natively inside Cursor, Warp, and every IDE at the system level.

4x faster than typing. 89% of messages sent with zero edits. Used by engineering teams at OpenAI, Vercel, and Clay.

Try Wispr Flow free

Today we’re kicking it off with a bit of a weird one… yes, the LinkedIn AI slop button 😏. 

Forward thinkers, welcome to issue #169 of the Neural Frontier! 

In this week’s issue, we’re unpacking interesting (maybe even a bit weird) updates from LinkedIn and Anthropic. 

Spoiler alert: there’s a new Claude model out there, and it has your name written on it 😅. 

In a rush? Here's your quick byte: 

🗑️ LinkedIn adds a “seems like AI slop” button!

🚀 Anthropic launches Claude Opus 5.

🧑‍💻 Anthropic’s AI models breached three companies during security tests!

⚡ The Neural Frontier’s weekly spotlight: 3 AI tools making the rounds this week.

🗑️ LinkedIn adds a “seems like AI slop” button!

LinkedIn is finally acknowledging what many users have complained about for months: the platform is drowning in low-quality, obviously AI-generated posts.

The company has introduced a new reporting option that lets users flag content that “Seems like AI slop.”

The button appears inside the three-dot menu on posts. Once selected, LinkedIn hides the post and thanks the user for the feedback.

It is part of a broader effort to make the platform feel less automated and more human again.

🤖 LinkedIn has an AI-content problem

AI-generated posts have become increasingly common on LinkedIn, particularly among users trying to publish more frequently or manufacture thought leadership at scale.

A recent analysis by AI detection company Pangram reportedly found that 41% of long-form LinkedIn posts appeared to be entirely AI-generated. Of course, using AI does not automatically make a post bad.

The real problem is the wave of generic, repetitive content that follows the same patterns: predictable hooks, forced vulnerability, excessive line breaks, vague business lessons, and conclusions that say very little.

LinkedIn Chief Product Officer Hari Srinivasan says reducing AI slop is now a major priority for the company. Alongside the new reporting button, LinkedIn is introducing classifiers designed to identify content that is:

• Likely to be low-quality AI slop.

• Repetitive, unoriginal, or generally low-value.

• Coming from outside a user’s network and unlikely to be useful.

Posts identified by these systems may appear less frequently in recommendations and suggested content.

🧠 Users will help train the feed

The new report button is not simply for removing individual posts. LinkedIn says feedback from users will help it improve the systems that determine what appears in people’s feeds.

The platform is also testing a feature that could inform creators when other users find their posts inauthentic or overly dependent on AI. That could be a more direct intervention than simply reducing a post’s reach, although LinkedIn has not explained exactly what this feedback will look like or whether it will affect distribution.

The company is also removing an AI tool that rewrites or “enhances” posts. It plans to replace it with a simpler proofreading feature that corrects errors without substantially changing the author’s voice.

🚀 Anthropic launches Claude Opus 5.

Anthropic has launched Claude Opus 5, a new model designed to deliver near-frontier performance without the frontier-level price tag.

The company says Opus 5 comes close to its more powerful Claude Fable 5 model on coding and knowledge work, while costing roughly half as much per task. It is also significantly more capable than its predecessor, Opus 4.8, despite retaining the same API pricing.

Anthropic is positioning Opus 5 as the model people can use every day: powerful enough for difficult professional work, but efficient enough to run regularly without burning through tokens.

It is now the default model for Claude Max subscribers and the strongest model available on the Claude Pro plan.

💻 Built for coding and complex knowledge work

Anthropic claims Opus 5 is now the leading model on several coding, automation, computer-use, and problem-solving evaluations.

Its biggest strength appears to be completing complicated tasks thoroughly rather than producing a quick first attempt and stopping there.

According to Anthropic, Opus 5 is better at:

• Identifying the underlying cause of software bugs.

• Verifying whether its own solution actually works.

• Recovering when the obvious approach fails.

• Creating additional tools or tests when none are available.

• Iterating on a task until it reaches a usable result.

On Frontier-Bench, which measures performance on difficult software-engineering tasks, Opus 5 reportedly more than doubled Opus 4.8’s performance while completing tasks at a lower average cost.

On CursorBench, it came within 0.5% of Claude Fable 5’s best score while costing roughly half as much per task.

The model also performed strongly on business automation and computer-use tests. Anthropic says that, even at its lowest effort setting, Opus 5 completed more end-to-end automation tasks than any competing model in its evaluation.

🛠️ Opus 5 is more proactive when things go wrong

Anthropic shared several examples intended to show how Opus 5 behaves differently from earlier models. In one test, the model was given an image of a machine component and asked to recreate it as a 3D model. However, it was deliberately prevented from viewing the image normally.

Rather than giving up, Opus 5 reportedly built its own computer-vision pipeline to extract the geometry from the raw image data, then used that information to reconstruct the part.

In another example, the model was asked to fix a real bug in an open-source package manager. It found the underlying cause and addressed an edge case that had been missed in the software community’s original patch.

🛡️ Anthropic says it is also its most aligned model

Alongside the performance gains, Anthropic says Opus 5 is its most aligned model so far. During internal testing, it reportedly showed lower rates of deceptive or reckless behaviour than Opus 4.8, Sonnet 5, and Fable 5. It was also less susceptible to being manipulated into completing harmful tasks.

Anthropic says it deliberately avoided training Opus 5 heavily on offensive cybersecurity tasks. Its safeguards will generally allow the model to inspect source code and identify vulnerabilities, while restricting activities such as:

• Generating working exploits.

• Conducting penetration tests.

• Scanning software binaries for weaknesses.

• Completing other high-risk offensive security tasks.

The restrictions are expected to trigger much less frequently than those on Fable 5, making Opus 5 more practical for legitimate security work.

🧑‍💻 Anthropic’s AI models breached three companies during security tests!

Anthropic has disclosed three incidents in which Claude gained unauthorized access to real company systems during cybersecurity evaluations.

The models were supposed to operate inside isolated test environments. Instead, a misconfiguration left an internet connection open, allowing Claude to reach live production systems belonging to three organizations.

The incidents were uncovered after Anthropic reviewed more than 141,000 evaluation runs following OpenAI’s recent breach of Hugging Face.

🧠 The models knew something was wrong

The incidents involved Claude Opus 4.7, Claude Mythos 5, and an unreleased internal research model.

All three had been told they had no internet access, which led some of them to assume the real systems they encountered were still part of the simulation. Their responses differed:

• Opus 4.7 recognized it was inside real systems but continued anyway, retrieving credentials and accessing production data.

• Mythos 5 noticed similar warning signs but convinced itself that the environment was simulated, then published a malicious package to the public PyPI registry.

• Anthropic’s newer research model stopped once it concluded the target was real.

Anthropic says it found no evidence that the models were pursuing goals of their own. They were simply trying to complete the tasks they had been assigned.

🛡️ Anthropic says normal safeguards were disabled

The models were tested without the safety monitoring and classifiers used in Claude’s public products. Anthropic says those protections would likely have blocked the behavior, but they were intentionally removed to measure the models’ raw cybersecurity capabilities.

The company is now tightening how internet access is configured and monitored during evaluations. It is also working with the independent research group METR to review the incidents.

Anthropic noted that two of the affected organizations had not detected the activity before being contacted.

⚡ The Neural Frontier’s weekly spotlight: 3 AI tools making the rounds this week.

1. 🎙️ ElevenLabs is an AI audio platform offering ultra-realistic text-to-speech in 70+ languages, voice cloning, AI music generation, sound effects, conversational AI agents, and speech-to-text transcription.

2. 🎬 Wistia Remix is an agentic video editor that turns webinars, podcasts, screen recordings, and other raw footage into polished, multi-source videos you can direct and refine through simple chat prompts.

3. 🚀 Pazi is an AI business-building platform that turns your idea into a personalized roadmap, assembles an AI team to execute tasks, and keeps projects, opportunities, and progress organized in one dashboard.

A weird week, honestly…

LinkedIn is trying to clean up the AI-generated fluff on its feed. Anthropic has a new model that’s cheaper to run and better at complex work. And then there’s the not-so-small matter of Claude breaking into real company systems during security tests.

So, a bit of progress, a bit of platform cleanup, and a pretty loud reminder that testing powerful models is not always neat.

Catch you next week on The Neural Frontier 👋!

PS: If you’re the “friend” this mail was forwarded to, and you enjoyed it, hit the Subscribe button to see more content like this every week 🙂.