Anthropic’s Opus 5 Is Better at Resisting Prompt Injection

Schneier on Security

July 31, 2026 at 5:23 PM

Overview

Anthropic's latest AI model, Opus 5, shows significant improvements in resisting prompt injection attacks compared to its predecessor, Opus 4.8. Researchers found that the likelihood of an attacker succeeding in manipulating the model dropped from 5.5% to 2.0% over 15 attempts. This makes Opus 5 the most secure model in the evaluation, outperforming all other non-Claude models, including Muse Spark, which had a success rate of 16.5%. The analysis also revealed that variants of GPT 5.6 were much more vulnerable, with the most capable version, Sol, having a 20% chance of being attacked successfully within the same number of attempts. This research is crucial as it demonstrates the ongoing challenges of securing AI models against targeted attacks, affecting developers and users who rely on these technologies for safe interactions.

Key Takeaways

Affected Systems: Anthropic Opus 5, Opus 4.8, Muse Spark, GPT 5.6 variants (Sol, Terra, Luna)
Timeline: Newly disclosed

Original Article Summary

The chart is interesting. On the IPI benchmark, Opus 5 improved over Opus 4.8, reducing the probability of an attacker succeeding within 15 attempts from 5.5% to 2.0%, and from 0.5% to 0.2% on 1 attempt. It also improved on Sonnet 5 (5.9% at k=15) and Mythos 5 (2.6%), making it the most robust model evaluated. Opus 5 also outperformed all non-Claude models on this benchmark. The most robust non-Claude model was Muse Spark at 16.5% within 15 attempts—more than eight times Opus 5’s rate. The most capable GPT 5.6 variant, Sol, was comparable to its predecessor GPT 5.5 (20.0% versus 20.8% within 15 attempts), and was 10 times as likely to be successfully attacked as Claude Opus 5 at 2.0%. The other GPT 5.6 variants are less robust, at 30.4% (Terra) and 43.9% (Luna). A single attempt against GPT 5.6 Sol succeeded 3.1% of the time, higher than the 2.0% an attacker achieved against Opus 5 after fifteen attempts...

Impact

Anthropic Opus 5, Opus 4.8, Muse Spark, GPT 5.6 variants (Sol, Terra, Luna)

Exploitation Status

No active exploitation has been reported at this time. However, organizations should still apply patches promptly as proof-of-concept code may exist.

Timeline

Newly disclosed

Remediation

Not specified

Additional Information

This threat intelligence is aggregated from trusted cybersecurity sources. For the most up-to-date information, technical details, and official vendor guidance, please refer to the original article linked below.

Read Original Article

Related Coverage

Trump blames Minnesota for cyberattacks on water sector, drawing pushback from cyber world

CyberScoop

In a recent statement, President Trump shifted the blame for a series of cyberattacks targeting the water sector away from Iran, which U.S. intelligence agencies have identified as a likely culprit. Instead, he pointed to Minnesota, implying that local authorities might be responsible for the incidents. This assertion has drawn significant criticism from cybersecurity experts who argue that it undermines the serious nature of the threats posed by foreign actors. The cyberattacks have raised alarms about the security of critical infrastructure, with potential implications for public safety and national security. Experts emphasize the need for a unified response to such threats, rather than attributing blame to local entities without evidence.

Jul 31, 2026

Suspected Chinese-Speaking Hackers Target Central Asian Governments With OctLurk and SilkLurk

The Hacker News

A group of Chinese-speaking hackers is reportedly targeting government entities in Central Asia, including countries like Afghanistan, Kyrgyzstan, Tajikistan, Uzbekistan, Kazakhstan, and Syria. Since January 2025, these attacks have focused on various sectors, particularly healthcare, research, and government offices. The cybercriminals are using tools known as OctLurk and SilkLurk to execute their operations. This rise in attacks poses significant risks to the affected governments and could compromise sensitive information and public services. As these countries navigate their cybersecurity challenges, the situation underscores the need for enhanced security measures to protect critical infrastructure.

Jul 31, 2026

Unspecified threat actors targeting US water systems, CISA warns

SCM feed for Latest

The Cybersecurity and Infrastructure Security Agency (CISA) has issued a warning about a rise in cyberattacks targeting programmable logic controllers (PLCs) used in U.S. water utilities. These PLCs are crucial for managing water treatment and distribution systems, making them a potential target for malicious actors. The increase in attacks raises concerns about the safety and security of drinking water and public health, as unauthorized access to these systems could disrupt services or lead to contamination. CISA advises water utilities to enhance their cybersecurity measures and remain vigilant against potential threats. The agency is also encouraging organizations to report any suspicious activity to help mitigate risks.

Jul 31, 2026

CISA Issues Fresh SBOM Guidance. Did They Get It Right?

darkreading

The Cybersecurity and Infrastructure Security Agency (CISA) has released updated guidance on Software Bill of Materials (SBOM), introducing a series of changes aimed at making the fields more thorough. While these updates are designed to enhance the documentation of software components, some experts believe that the revisions do not significantly address risk management concerns. Critics argue that without real improvements in how risks are assessed and managed, the changes may fall short in helping organizations better understand their software supply chain vulnerabilities. This guidance affects software developers and organizations that rely on third-party components, emphasizing the need for clearer documentation as cybersecurity threats continue to evolve. Companies should review the new guidance to ensure compliance and improve their security posture.

Jul 31, 2026

Hacker uses DeepSeek AI to autonomously attack vulnerable servers

BleepingComputer

A Chinese-speaking hacker is leveraging the DeepSeek AI model along with the open-source Hermes Agent to autonomously target vulnerable servers. This approach allows the attacker to conduct cyberattacks with minimal human input, increasing the efficiency and speed of these operations. The attacks focus on servers that are inadequately protected, which raises concerns for organizations relying on such systems. With the rise of AI-assisted hacking tools, the threat landscape for exposed servers is evolving, making it crucial for companies to enhance their security measures. This incident highlights the need for better server protection and continuous monitoring to prevent unauthorized access.

Jul 31, 2026

CISA warns of cyberattacks disrupting U.S. water utilities

BleepingComputer

The U.S. Cybersecurity and Infrastructure Security Agency (CISA) has issued a warning about a rise in cyberattacks targeting programmable logic controllers (PLCs) used in water and wastewater systems. These attacks are primarily aimed at systems that are connected to the internet, potentially allowing attackers to disrupt essential water services. The increase in these cyber threats raises significant concerns about the safety and reliability of public water supplies, as any successful intrusion could lead to harmful consequences for communities. CISA's alert emphasizes the need for water utilities to bolster their cybersecurity measures to protect against these vulnerabilities. This situation not only affects the utilities but also puts public health at risk if water services are compromised.

Jul 31, 2026