RESEARCH
May 2025: Rowhammer against post-quantum, TEEs bypassed, weak password managers
May 2025 is IEEE S&P month and hardware is back at the centre. Rowhammer extracts Kyber, BIKE and Dilithium keys, brief physical tampering with RAM breaks AMD SEV-SNP's guarantees and Apple chips speculate on data never read. On AI, anti-deepfake watermarks can be erased. 38 of 43 password managers remain exposed to offline attacks on the master password.
Archive: reconstruction written in November 2026
1. PQ-Hammer: End-to-End Key Recovery Attacks on Post-Quantum Cryptography Using Rowhammer
IEEE Symposium on Security and Privacy · Samy Amer, Yingchen Wang, Hunter Kippen, Thinh Dang et al. · Post-Quantum Cryptography
In short: Rowhammer fully recovers Kyber, BIKE and Dilithium secret keys, no supercomputer needed.
As post-quantum cryptography reaches standardisation, implementation security matters. PQ-Hammer combines Rowhammer, performance degradation and memory massaging to induce faults and fully recover the secret keys of Kyber, BIKE and Dilithium, with moderate effort and no months of precomputation. The authors recommend implementation techniques that harden the schemes.
Why it matters. Post-quantum migration does not remove platform risk: on shared servers a key can fall to a memory defect, not an algorithmic one.
2. BadRAM: Practical Memory Aliasing Attacks on Trusted Execution Environments
IEEE Symposium on Security and Privacy · Jesse De Meulemeester, Luca Wilke, David Oswald, Thomas Eisenbarth et al. · Hardware & Side Channels
In short: Brief tampering with RAM's SPD chip breaks AMD SEV-SNP's integrity guarantees and attestation.
Trusted execution environments now protect whole virtual machines with memory encryption. BadRAM shows that an attacker with brief physical access to the SPD chip on DDR4 and DDR5 modules can create aliases in the address space using cheap equipment. This bypasses CPU access controls and compromises SEV-SNP attestation. Scalable SGX and TDX block the attack.
Why it matters. Confidential computing assumes the cloud provider's hardware is untampered: anyone trusting it with sensitive data must weigh insider risk in the data centre.
3. SLAP: Data Speculation Attacks via Load Address Prediction on Apple Silicon
IEEE Symposium on Security and Privacy · Jason Kim, Daniel Genkin, Yuval Yarom · Hardware & Side Channels
In short: Apple chips' address predictor loads never-read data and lets attackers steal it, including from Safari.
Spectre-style attacks usually abuse control flow. SLAP instead exploits a data optimisation: the Load Address Predictor in Apple chips from M2 and A15 onwards preloads a predicted address that may point to secrets never read, with a wide speculation window. The authors read out of bounds, break ASLR and compromise the Safari browser.
Why it matters. Millions of Macs and iPhones share the same microarchitecture: a flaw like this creates widespread exposure that depends on vendor updates.
4. Security Analysis of Master-Password-Protected Password Management Protocols
IEEE Symposium on Security and Privacy · Yihe Duan, Ding Wang, Yanduo Fu · Identity & Authentication
In short: 38 of 43 password managers are exposed to at least one type of offline attack on the master password.
In password managers the client authenticates with the master password and the server helps retrieve credentials across devices, ideally knowing neither. The first formal analysis of these protocols, reconstructed from 43 managers through documentation, traffic analysis and reverse engineering, classifies four types of offline guessing attack: 38 managers fall to at least one.
Why it matters. After a provider breach, credential security rests on the master password alone: a textbook case of knock-on loss across many customers.
5. 403 Forbidden? Ethically Evaluating Broken Access Control in the Wild
IEEE Symposium on Security and Privacy · Saiid El Hajj Chehade, Florian Hantke, Ben Stock · Identity & Authentication
In short: An ethical way to test access control on live sites finds 19 exploitable flaws on 7 of 100 sites.
Broken access control is the most common web vulnerability per OWASP, but studying it on live sites risks exposing unwitting users' data. The Variable Swapping Framework uses two accounts per site and swaps identifiers between them, so only its own data are touched. Across 100 web apps it finds 584 sensitive endpoints and 19 exploitable flaws on 7 sites, all disclosed.
Why it matters. Access control underlies many personal data leaks: a safe test on production systems also helps in assessing suppliers.
6. UnMarker: A Universal Attack on Defensive Image Watermarking
IEEE Symposium on Security and Privacy · Andre Kassis, Urs Hengartner · AI & LLM Security
In short: UnMarker erases defensive watermarks from generated images, semantic ones included, without knowing the scheme.
Watermarks hidden in generated images are meant to flag deepfakes. UnMarker is the first practical universal attack: it needs no detector feedback or knowledge of the scheme, and disrupts the spectral amplitudes where robust watermarks must live. It defeats traditional schemes and cuts the best detection rate of semantic watermarks to 43%. Watermarking is no viable deepfake defence.
Why it matters. Anti-fraud checks on images and documents cannot rely on watermarks: controls independent of content are needed.
7. BAIT: Large Language Model Backdoor Scanning by Inverting Attack Target
IEEE Symposium on Security and Privacy · Guangyu Shen, Siyuan Cheng, Zhuo Zhang, Guanhong Tao et al. · AI & LLM Security
In short: BAIT finds whether an LLM hides a backdoor by reconstructing the target response, even with black-box access only.
LLM backdoors link a secret trigger to an attacker-chosen response, but the size of the output space defeats classic detection techniques. BAIT exploits the strong causal links among target-response tokens and reconstructs the target instead of searching for the trigger. It works without knowing trigger or target, even with black-box access only, and is evaluated on 153 LLMs.
Why it matters. Adopters of third-party models need an intake check: a backdoor scan is a concrete test before going into production.
8. GPTracker: A Large-Scale Measurement of Misused GPTs
IEEE Symposium on Security and Privacy · Xinyue Shen, Shen Yun, Michael Backes A, Yang Zhang · Malware & Threat Intelligence
In short: Of 755,297 custom GPTs, 2,051 serve forbidden purposes and 446 link to malicious domains.
Over eight months GPTracker collected 755,297 custom GPTs from the official store and 28,464 conversations. Combining LLM-based scoring with human review, it identifies 2,051 GPTs misused in ten forbidden scenarios. Builders evade review through external APIs, intent hidden in descriptions and redirects; 446 GPTs show 50 malicious domains, including phishing and malware.
Why it matters. AI agent stores are becoming a phishing and malware channel just like app stores: company AI-use policies need to cover them.
9. From Control to Chaos: A Comprehensive Formal Analysis of 5G's Access Control
IEEE Symposium on Security and Privacy · Mujtahid Akon, Md. Toufikuzzaman, Syed Rafiul Hussain · Network Security
In short: Formal analysis of 5G core access control finds five new classes of privilege escalation in the standards.
CoreScan builds the first full formal model of access control in the 5G core, including indirect communication and roaming, and verifies it by splitting it into components. Across 61 properties it finds five new classes of exploitable privilege escalation in the 5G standards, and shows most known over-privilege flaws extend to indirect communication and roaming.
Why it matters. Flaws in standards replicate across every operator adopting them: a shared-infrastructure risk with potential large-scale outages.
10. Mind the Location Leakage in LEO Direct-to-Cell Satellite Networks
IEEE Symposium on Security and Privacy · Weisen Liu, Zeqi Lai, Qian Wu, Hewu Li et al. · CPS & IoT Security
In short: Listening to direct-to-cell satellite signals locates active users to within a few hundred metres.
Low-orbit satellites that connect straight to phones are spreading. DCATOR monitors broadcast signalling messages, extracts location clues and combines them with satellite trajectories. On Iridium, the developing Starlink direct-to-cell service and a 3GPP NTN-based network, in the worst case it locates users within a few hundred metres.
Why it matters. Satellite connectivity is reaching ordinary phones: location leakage concerns anyone using it in sensitive areas or emergencies.
Trends of the month
Hardware & Side Channels
45 papers this month on this topic
IEEE S&P puts hardware back at the centre of risk. Rowhammer breaks new post-quantum signatures and encryption, a few minutes of RAM tampering defeats AMD's TEE guarantees, and an Apple chip optimisation opens a new family of speculative attacks. The common thread is that correct software is not enough: security depends on the platform it runs on.
Papers cited:
- PQ-Hammer: End-to-End Key Recovery Attacks on Post-Quantum Cryptography Using Rowhammer
- BadRAM: Practical Memory Aliasing Attacks on Trusted Execution Environments
- SLAP: Data Speculation Attacks via Load Address Prediction on Apple Silicon
AI & LLM Security
373 papers this month on this topic
With 373 of 2,053 papers, AI security is again the top topic in May. The papers chosen concern trust in models and their outputs: LLM backdoors can be found without knowing the trigger, anti-deepfake watermarks can be erased, and thousands of custom GPTs serve forbidden purposes, sometimes linking to phishing and malware.
Papers cited:
- BAIT: Large Language Model Backdoor Scanning by Inverting Attack Target
- UnMarker: A Universal Attack on Defensive Image Watermarking
- GPTracker: A Large-Scale Measurement of Misused GPTs
Identity & Authentication
70 papers this month on this topic
Two papers measure how well identity defences hold in real systems. Almost every password manager analysed leaves the master password open to offline attack if the server is compromised, and an ethical test of 100 web apps finds exploitable access-control flaws in 7. These are the weaknesses that turn one incident into losses for many customers.
Papers cited:
- Security Analysis of Master-Password-Protected Password Management Protocols
- 403 Forbidden? Ethically Evaluating Broken Access Control in the Wild
How this digest is made
This issue belongs to the digest archive, reconstructed after the fact. The month's papers were collected from arXiv, the IACR ePrint archive and the main security conferences and journals; the selection also weighs the citations received since publication. An AI model assessed the candidates and wrote the summaries, which are based on the abstracts: the original paper is the reference for details.
Papers of the month assessed: 2053.