Connect with us

CyberSecurity

Top AI Agents Built to Catch Malicious Code Can Be Tricked Into Running It

Published

on

AI coding agents

The AI Agents Meant to Protect Us Can Be Turned Against Us

Ask an AI coding agent to scan open-source code for security holes, and it might run the attacker’s code on your own machine instead. That’s the unsettling finding from a new proof-of-concept published Wednesday by the AI Now Institute. The attack, dubbed “Friendly Fire,” targets Anthropic‘s Claude Code and OpenAI‘s Codex — two of the most widely used AI coding agents — when either is running in an autonomous mode that approves its own actions.

The core issue isn’t that these tools are poorly built. It’s that they’re too trusting — of the code they analyze and of themselves. When a developer asks an AI coding agent to review a third-party library or a pull request for vulnerabilities, the agent might execute commands embedded in the very code it’s supposed to inspect. Attackers can craft malicious snippets that look like harmless code but contain hidden instructions for the AI to run.

How the Friendly Fire Attack Works

The attack exploits a fundamental design flaw in autonomous coding agents. These agents operate by reading code, analyzing it for security issues, and then — crucially — executing actions based on that analysis. In autonomous mode, they don’t stop to ask a human for permission before running a command. So if the code under review contains a carefully disguised instruction like a shell command or a hidden API call, the agent runs it without a second thought.

The AI Now Institute demonstrated this with Claude Code and Codex. In their tests, they embedded malicious payloads inside seemingly benign open-source packages. When the agents scanned these packages for vulnerabilities, they triggered the payloads — effectively compromising their own host environment. The attack doesn’t require sophisticated social engineering. It’s a simple bait-and-switch: hide a dangerous command inside code that looks safe to review.

“The agent doesn’t distinguish between code it should analyze and code it should execute,” the researchers noted in their paper. “It treats all code as equally trustworthy.”

Why Autonomous Mode Is the Problem

Autonomous mode is a feature, not a bug — at least from the developer’s perspective. It lets Claude Code and Codex work through long tasks without constant hand-holding. You ask it to audit a codebase, and it plows through hundreds of files, flagging issues and even suggesting fixes. That speed is what makes these tools valuable.

But speed comes at a cost. In autonomous mode, the agent approves its own actions. It doesn’t ask, “Hey, should I run this command?” It just runs it. That’s fine when the code is clean. But when the code is malicious, the agent becomes an unwitting accomplice. The attacker doesn’t need to break into your system; they just need to trick the tool you trust into doing it for them.

The researchers point out that this isn’t a theoretical risk. Open-source code is the backbone of modern software development, and developers regularly use AI coding agents to scan it for flaws. If an attacker can poison a popular package — or even a single pull request — with a payload that targets these agents, they can compromise every machine that scans it.

Both Anthropic and OpenAI Have Been Notified

The AI Now Institute says it responsibly disclosed the vulnerability to both Anthropic and OpenAI before publishing the proof-of-concept. As of now, neither company has released a patch or a public statement addressing the specific attack. The researchers note that the issue isn’t limited to these two tools — any AI coding agent running in an autonomous mode that executes code from the input could be vulnerable.

Claude Code and Codex are the most prominent examples, but the attack pattern likely applies to other agents that follow similar design principles. The researchers recommend that developers avoid using autonomous mode when scanning untrusted code. If you must use it, they suggest running the agent in a sandboxed environment with no network access and limited file system permissions.

“The safest approach is to never let an AI agent execute code it’s analyzing,” the paper states. “Treat all external code as hostile until proven otherwise.”

What Developers Can Do Right Now

Until the companies roll out fixes, developers have a few options to protect themselves:

  • Disable autonomous mode when scanning code from unknown sources. Require explicit human approval for every action.
  • Run AI coding agents in isolated environments — containers, virtual machines, or sandboxes with restricted permissions.
  • Audit the agent’s actions manually before approving them. Don’t trust the agent’s judgment blindly.
  • Use static analysis tools that don’t execute code as part of their workflow. Traditional linters and vulnerability scanners don’t have this problem.

None of these are perfect solutions. Disabling autonomous mode defeats the purpose of using AI agents for speed. Running them in sandboxes adds overhead. But for now, it’s the best defense available. The Friendly Fire attack is a reminder that AI security flaws aren’t just about models hallucinating or leaking data — they can also turn your tools into weapons.

The broader lesson is that AI coding agents need a fundamental redesign. They should treat code as data, not as commands. They should never execute code they’re analyzing unless explicitly instructed to — and even then, only with human oversight. Until that happens, every developer using Claude Code or Codex in autonomous mode is one malicious pull request away from a compromised machine.

“We built these tools to help us write safer code,” the researchers conclude. “But we forgot to make them safe themselves.”

Continue Reading

CyberSecurity

PostgreSQL Patches 12-Year-Old Logical Decoding Flaw That Allowed Code Execution

Published

on

logical decoding flaw

PostgreSQL Ships Emergency Updates for a Decade-Old Security Hole

The PostgreSQL Global Development Group has pushed out urgent patch releases to close a security vulnerability that lets anyone with the REPLICATION attribute run arbitrary code as the database server’s operating-system user. The bug, tracked as CVE-2026-6471, carries a CVSS score of 7.2 — high severity by any measure.

What makes this one particularly nasty is its age. The flaw has been lurking since logical decoding was first introduced in PostgreSQL 9.4 back in 2014. That’s twelve years of exposure. Twelve years of potential exploitation for anyone who managed to obtain replication privileges.

The fix lands in versions 18.6, 17.11, 16.15, 15.19, and 14.24. If you’re running any earlier release, you’re vulnerable. No ifs, ands, or bugs.

What Exactly Is Logical Decoding, and Why Should You Care?

Logical decoding is a feature that lets you extract changes from a PostgreSQL database in a format that’s independent of the physical storage. It’s what powers streaming replication, change data capture (CDC) pipelines, and many modern data integration tools. In short, it’s the backbone of real-time data movement for countless organizations.

But here’s the catch: the vulnerability allows a user with the REPLICATION attribute — a role typically reserved for backup and replication processes — to escalate privileges and execute code as the OS user running PostgreSQL. That means an attacker who compromises a replication account could potentially take over the entire database server, read sensitive files, or even pivot to other systems on the network.

Who’s at Risk?

Any organization running PostgreSQL with logical decoding enabled and replication roles granted to non-trusted users is in the danger zone. Even if you don’t use logical decoding actively, the vulnerability exists in the code path, and a determined attacker could trigger it.

The PostgreSQL team’s advisory is clear: upgrade immediately. There are no workarounds that fully mitigate the issue, though restricting REPLICATION privileges to only the most trusted accounts can reduce your attack surface.

A Timeline of Neglect: How a 12-Year-Old Bug Survived

It’s almost unbelievable that a flaw this severe could persist for over a decade. Logical decoding was a major feature addition in 9.4, and it’s been a core part of PostgreSQL’s appeal ever since. Yet somewhere in the complex code that handles replication slots and WAL (write-ahead log) processing, a subtle bug slipped through.

Security researchers have long noted that PostgreSQL’s security record is generally solid — but this incident is a stark reminder that even the most reputable open-source projects can carry hidden landmines. The fact that it took this long to discover highlights the challenges of auditing complex, long-lived codebases.

What Should PostgreSQL Admins Do Right Now?

Here’s a practical checklist for anyone running PostgreSQL:

  • Identify your current version: SELECT version(); or check with your package manager.
  • If you’re on 14.x, 15.x, 16.x, 17.x, or 18.x, upgrade to the latest patch release listed above.
  • If you’re on an older version (e.g., 13 or below), you need to plan a major upgrade — those branches are no longer supported and won’t receive fixes.
  • Audit all roles that have the REPLICATION attribute. Revoke it from any account that doesn’t absolutely need it.
  • Review your database logs for any suspicious activity related to logical decoding or replication slots.

Don’t Forget About Your Replication Setup

If you’re using streaming replication or a tool like Debezium that relies on logical decoding, pay extra attention. Your replication slots might be active, and the vulnerability could be triggered through them. After upgrading, test your replication thoroughly to ensure nothing breaks.

Also, consider using a connection pooler or proxy to limit direct database access. It’s not a fix for this specific bug, but it’s good defense-in-depth practice.

Broader Implications for Open-Source Security

This incident raises uncomfortable questions about the sustainability of security auditing in open-source projects. PostgreSQL is maintained by a dedicated community, but it’s a massive codebase. Finding a bug that’s been hidden for 12 years requires either luck or a very thorough review.

For organizations that rely on PostgreSQL — and that’s a huge portion of the internet’s data infrastructure — this is a wake-up call. Regular security audits, prompt patching, and a strong understanding of your database’s privilege model are non-negotiable.

If you’re also using tools that interact with PostgreSQL’s logical decoding, like change data capture tools, make sure those are updated as well. And if you’re new to PostgreSQL security, check out our PostgreSQL hardening guide for baseline practices. For broader context, see how database security advisories are handled across major systems.

The bottom line: don’t wait. The exploit is public knowledge now, and attackers will be scanning for vulnerable instances. Patch your systems, tighten your roles, and hope that this 12-year-old skeleton in PostgreSQL’s closet is the last one.

Continue Reading

CyberSecurity

Millions of Phishing Emails Hide ‘Funding’ in Invisible Unicode to Slip Past Filters

Published

on

invisible Unicode phishing

Microsoft Warns of High-Volume Phishing Campaign

Microsoft’s security team has issued an alert about a phishing campaign that’s been blasting out millions of emails. The trick? Invisible Unicode tag characters that split financial lure words like ‘funding’ to dodge email filters.

Instead of hiding instructions from humans while exposing them to AI models, the attackers used these invisible characters to break up keywords. That way, the filters never see the full word — but the recipient’s email client renders it seamlessly.

How the Invisible Unicode Attack Works

Unicode tag characters are normally used for language tagging in plain text. They’re invisible in most rendering engines. Attackers inserted them mid-word, so ‘funding’ becomes ‘f-u-n-d-i-n-g’ with invisible tags between each letter.

Filters that scan for exact strings don’t match. The email lands in the inbox looking perfectly normal. It’s a clever piece of social engineering that targets the gap between what machines parse and what humans read.

The Role of AI in Detection

Microsoft’s research team noted that while AI models can often spot these anomalies, the attackers deliberately avoided that route. They weren’t trying to trick AI — they wanted to slip past traditional signature-based filters that haven’t caught up to Unicode obfuscation.

This marks a shift in tactics. Earlier campaigns used Unicode to hide malicious instructions from human reviewers while keeping them visible to AI. This one flips the script entirely.

Why Financial Lure Words Matter

Words like ‘funding’, ‘transfer’, and ‘invoice’ are common hooks in business email compromise (BEC) scams. By splitting them, attackers ensure their emails don’t trigger the same automated checks that would normally flag them.

  • Filter evasion: Splitting keywords means regex patterns and string matches fail.
  • Human perception: Invisible characters don’t alter how the email looks to a recipient.
  • Scale: Microsoft describes this as high-volume, meaning millions of emails are involved.

The campaign appears to target organizations that handle financial transactions — payroll departments, accounts payable teams, and CFOs.

How to Protect Against Unicode Phishing

Email security teams need to update their detection rules. Look for emails that contain Unicode tag characters (U+E0000 to U+E007F) in suspicious positions, especially inside common financial keywords.

Regular users should be cautious of unexpected emails asking for wire transfers or payment changes, even if they look legitimate. Verify requests through a second channel — a phone call, not a reply email.

For more on protecting yourself, check out our guide on recognizing phishing email signs. If you’re dealing with a potential breach, our article on incident response steps for small businesses can help you react quickly.

Technical Mitigations

Administrators can configure their email gateways to either strip or flag Unicode tag characters in incoming messages. Microsoft Defender for Office 365 has also been updated to detect this pattern, but organizations using other filters should test their own systems.

Security researchers recommend adding decoy keywords to honeypots — traps that catch attackers when they use the same obfuscation technique.

The Bottom Line

This campaign is a reminder that email filters are only as good as their understanding of attacker tricks. Invisible Unicode is not new — but using it to split lure words for mass distribution is a fresh twist that many defenses aren’t ready for.

Stay alert. If an email asks for money or sensitive data, scrutinize it — even if it looks perfect. And if you’re in IT, audit your mail flow for Unicode anomalies before the next wave hits.

Continue Reading

CyberSecurity

Microsoft Cloud Patches, 5,000 Hacked Dropbox Accounts, and a $1.1B Security Startup: What You Missed

Published

on

Microsoft cloud patches

The Week’s Under-the-Radar Security Stories

Some stories don’t get the headline treatment they deserve. They still matter, though. This week’s quiet-but-significant batch includes a wave of cloud patches from Microsoft, a credential-stuffing attack on Dropbox, and a cybersecurity startup hitting unicorn status.

Here’s what you need to know.

Microsoft Rolls Out Patches for Cloud Services

Microsoft has been busy behind the scenes. The company pushed out fixes for several of its cloud offerings, addressing vulnerabilities that could have given attackers a foothold in enterprise environments.

The patches cover a range of services, though Microsoft hasn’t disclosed every detail. What’s clear is that IT teams should treat these updates as priority. Cloud misconfigurations and unpatched flaws remain a top attack vector, and this is a reminder that even the biggest providers need constant upkeep.

For admins, the takeaway is straightforward: check your Microsoft cloud security dashboard, review the latest advisories, and apply the updates before they become a problem. Delaying patches in a cloud environment is a gamble, and the house usually wins.

What the Patches Target

Microsoft’s advisory points to vulnerabilities in Azure and related services. Specifics are sparse, but the company’s track record suggests these could range from privilege escalation to information disclosure. If you’re running any Microsoft cloud workload, the official security update guide is your first stop.

5,000 Dropbox Accounts Hacked via Credential Stuffing

Dropbox confirmed that attackers compromised roughly 5,000 user accounts. The method? Credential stuffing — using usernames and passwords stolen from other breaches to break into accounts where people reuse passwords.

This isn’t a breach of Dropbox’s own systems. The company says its infrastructure wasn’t compromised. Instead, the attackers leveraged the all-too-common habit of password reuse. Once they had valid credentials from elsewhere, they simply tried them on Dropbox.

Dropbox has reset passwords for affected users and is rolling out additional protections. But the incident underscores a persistent problem: credential stuffing attacks remain one of the most effective ways for hackers to get in. No fancy exploits needed, just a list of leaked passwords and a bit of patience.

How to Protect Yourself

  • Use a unique password for every account. Yes, every single one.
  • Enable two-factor authentication, especially on cloud storage and email.
  • Check haveibeenpwned.com to see if your credentials have been exposed.
  • If you’re a Dropbox user, change your password now, even if you weren’t affected.

It’s tedious, but it works. The hackers who did this weren’t geniuses — they were just counting on people to make the same mistake twice.

Guardio Hits $1.1 Billion Valuation

In brighter news, Guardio, a browser security startup, has reached a valuation of $1.1 billion. The company, which focuses on protecting consumers from phishing, malware, and malicious extensions, has been growing quietly but steadily.

Guardio’s approach is simple: a lightweight browser extension that blocks threats before they reach the user. It’s a consumer-focused product, but the underlying tech has broader implications. As more people work from home, the browser has become the new perimeter.

The Guardio funding round signals that investors see value in endpoint protection that doesn’t require a degree in cybersecurity to operate. That’s a good sign for the industry, and an even better one for users who just want to browse without getting hacked.

Why These Stories Matter

On the surface, these three items seem disconnected. A cloud patch, a credential stuffing attack, and a funding round — what’s the thread?

It’s this: security is a moving target. Microsoft’s patches show that even the giants are constantly fixing holes. The Dropbox incident shows that human behavior — password reuse, ignored 2FA — often undoes even the best technical defenses. And Guardio’s valuation shows that the market rewards products that make security accessible.

None of these stories will dominate tomorrow’s headlines. But together, they paint a picture of an industry that’s always fighting, always adapting, and always finding new ways to protect users. That’s worth paying attention to, even if it doesn’t make the front page.

Stay patched, stay vigilant, and for heaven’s sake, stop reusing your passwords.

Continue Reading

Trending